Cerebras' Low-Latency Inference Bet: Can the OpenAI Contract, Delivery Economics, and Cash per Share Close the Loop?

Cerebras Systems (NASDAQ: CBRS) In-Depth Equity Research Report

Chapter 1: One-Page Conclusion

Cerebras' question is not "will AI compute grow?" That question has already become too large, and too easy to misjudge. The real question is whether this company can turn its speed advantage, the large OpenAI contract, and the inference services it still has to deliver into gross profit, cash flow, and real per-share value that common shareholders can actually receive.

The First Question Readers Should Ask	Current View
Where things stand now	The technology path, OpenAI contract, roughly $24.6 billion future revenue contract pool, and IPO pricing have all been seen by the market; the company has entered the validation stage for delivery, gross margin, cash flow, and the equity denominator.
One-sentence conclusion	Cerebras is not an ordinary AI chip company, and it is not a full replacement for Nvidia. It looks more like an AI infrastructure company that packages wafer-scale chips, dedicated systems, and high-speed inference services for sale.
Positive signs already visible	OpenAI has committed to purchase 750MW of inference service capacity, and the contract pool is very large; channel partnerships such as AWS, the inference interface, CS-3/WSE technical evidence, and workload clues from code agents, enterprise search, medical research, national labs, and real-time voice and video all show that speed does have application value.
Evidence still missing	Whether these commitments can become available services on schedule, whether the future revenue contract pool can become high-margin revenue, whether customers beyond OpenAI can be replicated, whether operating cash flow can improve, and whether warrants, options, RSUs, and lock-up supply will depress per-share value.
What would make things better	The future revenue contract pool is recognized on schedule, gross margin is stable, operating cash flow improves, inference interfaces and model services scale beyond OpenAI, and the equity denominator remains controlled.
What would make things worse	Delivery delays, gross margin below expectations, capital expenditures consuming cash, no replication beyond OpenAI, cloud platforms capturing the customer gateway, and lock-up releases plus warrant dilution weighing on per-share value.

Cerebras' technical highlights already exist, and the market has already assigned high expectations. The real main line is not "is the chip special enough," but rather: can the speed advantage become a deliverable service, can the deliverable service become revenue, can revenue become cash, and can enterprise value growth outpace expansion in the equity denominator?

The core view is this: Cerebras is a high-quality new issue worth following closely, but it is not a low-expectation new issue. Every future upward revision must come from harder evidence on delivery, gross margin, cash, and per-share value, not from a larger AI imagination space.

---

Chapter 2: How Inference Capacity Becomes Cash per Share

If Cerebras is treated as a company that "sells big chips," the numbers that follow quickly turn into a string of jargon. The smoother entry point is to first look at what customers are actually buying.

OpenAI or enterprise customers are not paying for a nice-sounding technical term. They are paying for a pool of inference capability that can go online, respond reliably, and help AI products complete tasks faster. What Cerebras has to prove is not that "the chip is special," but whether this inference capability can be delivered on time, used continuously, and leave gross profit and cash behind.

This can be broken into three sentences.

First, the roughly $24.6 billion future revenue contract pool is not money already received. It is more like a batch of signed service commitments: customers have agreed to buy services in the future, but Cerebras still has to prepare systems, data centers, power, cooling, and operations, and revenue will enter the financial statements only gradually after the services reach an available state.

Second, 750MW of capacity is not an abstract number. It means Cerebras must deliver inference service capability at data-center and power scale. The larger the number, the stronger the revenue visibility; at the same time, it also means heavier construction, acceptance, depreciation, cash consumption, and execution risk.

Third, low latency is not benchmark showmanship. It is user waiting time: how long it takes for AI to start answering, whether each step of a code agent stalls, and whether a voice assistant can converse like a real person. If customers are willing to pay for this experience, Cerebras may turn speed into revenue; if customers treat it only as temporary acceleration, speed will be hard to turn into long-term profit.

Phrase Appearing in the Report	First Translate It As	Where It Is Truly Useful	Cannot Assume
Future revenue contract pool	Service commitments signed but not yet recognized as revenue	Shows future revenue visibility, especially related to the OpenAI contract	It is not cash, not profit, and not successful delivery already completed
750MW capacity	Data-center-scale inference service capability Cerebras has to deliver	Shows that the customer is not asking for a small trial, but for large-scale available services	It is not revenue already received, and not every MW will necessarily leave high gross margin
Low-latency inference	Allows AI to start answering faster and complete multi-turn tasks faster	Important for code agents, voice and video, enterprise search, and real-time interaction	It does not mean all training, all inference, and all models must use Cerebras

Therefore, the right reading order for Cerebras is: why customers need faster inference, whether Cerebras can deliver by capacity, whether revenue can be recognized, whether gross margin can hold, whether cash can flow back, and finally whether that value truly reaches each share after warrants, options, and lock-up supply.

Judging Cerebras cannot rely only on signed contracts, and cannot rely only on chip speed. The real path is longer: customers first pay for speed and dedicated inference services; then the company turns contracts into usable data centers and systems, turns available services into revenue, turns revenue into gross profit, and only after deducting capital expenditures and equity dilution can it potentially become per-share value received by common shareholders.

Component	Meaning	Where to Look	Easiest Misread
Demand for speed	Why customers are willing to pay	Industry position, demand pool, customer workloads	Fast speed does not equal pricing power
OpenAI / future revenue contract pool	Revenue visibility and delivery commitments	OpenAI contract and remaining performance obligations	The contract pool is not cash and not profit
Product entry points	How customers buy, call, and distribute	Product-line economics	More entry points do not mean the profit belongs to Cerebras
Data center and system delivery	Whether the company can deliver usable services on schedule	Delivery economics, three financial statements	MW is not revenue already recognized
Gross profit and operating cash flow	Whether revenue leaves value behind	Financial quality, gaps between metrics	Revenue growth does not equal free cash flow
Equity denominator	How enterprise value reaches each share	Common-shareholder section	A larger company does not mean per-share value rises in sync
Current price	Which successes the market has prepaid	Valuation expectations	A good company does not equal a good price

This path reduces two misjudgments: first, the large AI compute market does not mean Cerebras can retain profit; second, low-latency technology is very distinctive, but it still cannot be treated in advance as free cash flow already realized.

---

Chapter 3: Why the Market Gets Excited First

Cerebras' IPO can easily trigger excitement not just because it is another AI company, but because it hits three market sentiment points at once: OpenAI, Nvidia's moat, and the AI new-issue window.

First, OpenAI's presence means Cerebras is no longer merely a "chip story with special technology." When the strongest model company is willing to sign a large-scale inference service commitment, the market naturally reads it as an infrastructure clue that may affect the speed of next-generation AI. That judgment is reasonable, but it cannot be overextended. OpenAI is strong validation; it does not mean all customers will replicate it, and it does not mean this contract has already become high-margin cash.

Second, Cerebras is easy to package as "a company challenging Nvidia." This claim has communication power because Nvidia's position in AI chips is so strong, and the market has been looking for a second path. But the more accurate statement is not that Cerebras will fully replace Nvidia, but that it is trying to provide a fast lane different from GPU clusters in low-latency inference, code agents, real-time interaction, and some dedicated model services. It challenges part of Nvidia's default path, not the entire Nvidia ecosystem.

Third, the AI IPO window itself amplifies the story. The market is expecting more listings from AI infrastructure, model companies, data companies, and space technology companies. As one of the few public-market new issues that can directly tell an "AI hardware + OpenAI + inference speed" story, Cerebras will naturally be treated as a representative name for this window. That attention can bring liquidity and a valuation premium, and it will also raise the validation threshold for future earnings reports.

Therefore, market excitement itself is not wrong. The mistake is treating excitement as evidence. What can truly enter valuation remains the following: whether capacity goes online, whether the contract pool is recognized as revenue, whether gross margin holds, whether operating cash flow improves, whether customers beyond OpenAI replicate, and whether the equity denominator remains controlled.

Why the Market Is Excited	What Is Reasonable	What It Does Not Prove
OpenAI attachment	A top-tier customer is willing to commit to large-scale inference services, showing the technology and delivery capability are not empty narrative	It does not prove non-OpenAI customers will replicate, and does not prove gross margin is high enough
Nvidia challenger narrative	The market needs alternatives outside the GPU default path, and Cerebras' wafer-scale architecture is differentiated enough	It does not prove Cerebras can fully replace the Nvidia ecosystem
AI IPO window	Scarce new issue, strong theme, low float, and high trading volume can reinforce one another	It does not prove the first-day price is already cheap, and it does not eliminate lock-up supply
Speed story is easy to understand	User wait time, code agents, and voice interaction can all feel the speed difference	It does not prove the speed premium will necessarily be retained by Cerebras

This is Cerebras' first-layer contradiction: it does have a strong enough story for the market to assign high expectations in advance; but precisely because the market has already assigned high expectations, every future quarter must fill in the evidence with harder data.

---

Chapter 4: Four Variables: Speed, OpenAI, Delivery, and Share Supply

The key now is not to find more narratives, but to put four variables on the same page. Only if they improve together can Cerebras upgrade from "a new issue with strong customer validation" to "an inference infrastructure company capable of continuously creating per-share value."

Variable	Current State	Positive Signal	Negative Signal
Speed	The technical differentiation is real, but applicable scenarios do not equal all AI inference	More real-time interaction, code agents, voice, and enterprise workflows are willing to pay	GPU / TPU / Trainium / Groq narrow the gap, and Cerebras can serve only a small number of models or contexts
OpenAI	750MW and the future revenue contract pool are strong validation, but also sources of customer concentration and complex terms	Services go live on schedule, payment and acceptance timing is clear, and non-OpenAI customers follow	The OpenAI special-project character deepens, and replication among other customers is insufficient
Delivery economics	Data centers, power, cooling, hardware, and operations are hard constraints before revenue	Gross margin is stable, operating cash flow improves, and capital intensity declines	Revenue grows, but costs, depreciation, working capital, and cash consumption expand in step
Share supply	First-day close of $311.07, with the market already assigning high expectations	Under high trading volume, the price can absorb early releases, the overallotment, and lock-up expirations	The stock falls on high volume around lock-up releases, or dilution grows faster than operating delivery

These four variables set the research rhythm for Cerebras. Speed explains why customers may pay; OpenAI explains why it is not a concept company; delivery economics explain whether revenue can be retained; share supply explains why a good company may still not be a good entry point.

---

Chapter 5: First Positioning: What Cerebras Buys Is Not the Entire AI Compute Market

The phrase AI compute is too broad. It includes Nvidia's general-purpose GPUs and full-rack systems, Google, AWS, and Microsoft's in-house chips; compute clouds such as CoreWeave; Broadcom and Marvell helping large customers build custom chips; and further down, memory, networking, power, liquid cooling, and data centers. Every layer is benefiting from AI, but each layer earns money differently, retains profit differently, and consumes cash at a different pace.

Therefore, Cerebras cannot be understood through the "entire AI compute market." The more accurate framing is this: it is competing for a narrower and more specific budget, namely inference capacity that model companies, cloud platforms, and enterprise customers are willing to pay for because it delivers faster responses, shorter waiting time, and more stable real-time interaction.

5.1 AI Compute Infrastructure First Splits Into Five Layers

What this value chain sells is not a single chip, but compute that customers can call reliably. The chip is only the starting point. What customers actually buy is whether models can run, interfaces can remain stable, tasks can be completed faster, and costs can be controlled.

Layer	Representative Players	What Customers Actually Buy	How Money Flows	Meaning for Cerebras
Customer workloads	OpenAI, Anthropic, Google, Meta, enterprises, governments, research institutions	Training, inference, code generation, voice and video, enterprise search, multi-step agent tasks	Budgets come from models, applications, cloud services, and enterprise IT	Determines whether low latency is truly mission-critical
Cloud and interface gateways	AWS, Azure, Google Cloud, CoreWeave, OpenRouter, Hugging Face, Cerebras, Groq	Directly callable models, compute capacity, managed services, and interfaces	The party that owns the customer gateway is usually more likely to retain profit	Cerebras must prove it is not just a compute backend hidden behind others
Software runtime environment	CUDA, ROCm, XLA, Neuron, Cerebras SDK, model-serving tools	Developer migration cost, scheduling, stability, deployment convenience	Software habits determine customer switching costs	Nvidia and cloud vendors have strong advantages, and Cerebras needs to lower the customer adoption threshold
Accelerators and systems	Nvidia, AMD, TPU, Trainium, Maia, Cerebras, Groq, custom ASIC	Chips, servers, full-rack systems, dedicated inference equipment	Hardware revenue appears first, but gross margin depends on supply, cost, and utilization	Cerebras directly participates in this layer, but cannot stop at hardware sales
Physical infrastructure	HBM, networking, optical modules, power, liquid cooling, data centers	Memory, interconnect, power capacity, heat dissipation, and facilities	These links determine delivery speed and cost	They are not Cerebras' main revenue source, but they will determine whether 750MW can truly go online

Cerebras' position is unusual: it is neither a pure chip company nor a pure cloud services company. It starts from wafer-scale chips and CS systems, sells inference capacity, interface services, and model services upward, and downward must also bear the pressure of data centers, power, cooling, inventory, depreciation, and operations. This position makes it closer to customer tasks than an ordinary chip supplier, and heavier than an asset-light software company.

5.2 Revenue, Profit, and Cash Are Not in the Same Layer

The easy misread in AI infrastructure is to treat "who gets revenue" as "who keeps profit," and then treat "who has contracts" as "who has already generated cash." These three things must be separated.

Economic Layer	Who Is More Likely to Get Revenue	Who Is More Likely to Retain Profit	Can Cerebras Participate Directly?	Investment Meaning
General-purpose GPUs and full-rack systems	Nvidia, AMD, server makers, cloud vendors	Nvidia and cloud platforms that control the customer procurement path	Mainly a competitive relationship, not Cerebras' primary battlefield	This is the default path, the largest in scale and the hardest to bypass
AI cloud and compute capacity	Cloud vendors, CoreWeave, Oracle Cloud, Lambda, Cerebras, etc.	Platforms with high utilization, low cost of capital, and the ability to retain customers long term	Highly relevant; the OpenAI contract is essentially a large-scale capacity commitment	This determines whether Cerebras can move from a hardware company to a service-type compute platform
Inference interfaces and model services	OpenAI, Anthropic, Google, cloud vendors, Cerebras, Groq, developer platforms	The party that owns the model gateway, customer gateway, and cost advantage	There is an opportunity, but usage and customer retention still need to be proven	This is the part that most resembles the future profit pool, and also the part with the least sufficient evidence
Custom chips and interconnect	Broadcom, Marvell, cloud vendors' in-house teams	Suppliers with strong design capability and deep project attachment, plus the large customers that ultimately control the platform	Indirectly relevant, and it can also divert large customers' in-house demand	The more successful non-GPU paths become, the larger Cerebras' opportunity; but customers also become more likely to bypass it
Memory, networking, power, liquid cooling, data centers	HBM vendors, network chips, optical modules, power equipment, data-center operators	Links with tight supply, difficult certification, and strong delivery capability	Mainly cost and delivery constraints	These links will affect Cerebras' gross margin, capital expenditures, and go-live schedule

The conclusion from this map is direct: what Cerebras should fight for most is not "all AI chip revenue," but the two layers of low-latency inference capacity and inference interfaces. If it only sells hardware, revenue may be large but cash quality may not be good; if it can repeatedly sell capacity to multiple customers and make the interface part of customers' daily invocation path, profit quality can improve significantly.

5.3 Which Layer Does Cerebras Actually Occupy?

Disaggregated, Cerebras has four identities at once. Each identity has a different valuation meaning and enters different financial statement lines.

Cerebras' Identity	Corresponding Product or Contract	What It Can Show	What It Still Cannot Show	Where It Ultimately Lands
Wafer-scale system company	WSE-3, CS-3 systems	The technology path is differentiated enough and suited to reducing some cross-card communication and latency	It does not automatically equal high gross margin, and does not automatically mean customers are willing to migrate at scale	Hardware revenue, inventory, fixed assets, depreciation, cost of sales
Inference capacity supplier	OpenAI 750MW inference service commitment	A top-tier customer is willing to sign a contract for large-scale dedicated capacity	The contract pool is not cash and not profit already recognized	Future revenue contract pool, revenue recognition, receivables, operating cash flow, capital expenditures
Inference interface and model services platform	Inference API, public models, Cerebras Code, model services	Customers can use Cerebras' speed through interfaces and products	Usage, retention, and non-OpenAI customer scale still need validation	Usage revenue, service gross margin, customer concentration
Compute backend in partner channels	AWS, Hugging Face, OpenRouter, Poe, Vercel, and other partner gateways	Distribution friction is reduced, and developers can more easily access Cerebras	Expanded channels do not mean the customer gateway and pricing power belong to Cerebras	Channel revenue, revenue-sharing arrangements, gross margin, and customer ownership

In one sentence: Cerebras' most valuable possibility may not be "making a very large chip," but turning that chip into a low-latency inference service customers are willing to call repeatedly. The former proves engineering capability; the latter is what may prove commercial quality.

5.4 Relationship With Nvidia

Cerebras is not "the next Nvidia." Nvidia is the default path for AI factories. When customers buy Nvidia, they buy not only GPUs, but also CUDA, NVLink, networking, libraries, developer tools, developer ecosystem, cloud-vendor procurement habits, and enterprise deployment paths. For most customers, this is the path of least resistance, lowest risk, and most familiar supply chain.

Cerebras' opportunity is not to fully replace this main road, but to provide a dedicated fast lane for some high-value inference tasks. Real-time voice, code agents, multi-step agents, long outputs, enterprise search, and low-latency model services all share one feature: user waiting time directly affects experience and usage frequency. In these tasks, speed itself may become a reason customers are willing to pay.

Comparison Dimension	Nvidia Default Path	Cerebras Dedicated Path	What It Means for Investors
Why customers choose it	Stable, familiar, complete ecosystem, low migration risk	Faster for some inference tasks, lower latency	Cerebras cannot rely only on benchmarks; it must prove customers are willing to pay for the experience difference
Software habits	CUDA and cloud platform toolchains are already deeply embedded	Still needs to reduce developer and enterprise migration friction	If a technical advantage is too troublesome to use, it will be absorbed by the default path
Hardware form factor	GPUs, networking, and full-rack systems scale together	Wafer-scale system that reduces some cross-device communication	Advantages are more likely to appear in specific tasks, not all AI compute
Economic outcome	Scale, software, and ecosystem jointly retain profit	Delivery cost, utilization, and gross margin still need validation	Speed must pass through revenue, gross margin, and cash before it matters for per-share value

Therefore, the relationship between Cerebras and Nvidia is not "who eliminates whom," but a competition between the default path and a dedicated path. The default path suits most customers. The dedicated path forms an independent budget only when the experience difference is large enough, costs are controllable, and customers are willing to use it long term.

5.5 Relationship With Cloud Vendors' In-House Chips

Google TPU, AWS Trainium, Microsoft Maia, and Meta MTIA show that large customers will not buy only external GPUs forever. As long as inference scale is large enough, customers will try to build in-house, customize, optimize costs, and keep more profit inside their own platforms. For Cerebras, this is both positive and pressure.

The positive is that the market has already accepted that "non-GPU paths" can exist. The pressure is that the largest and most workload-savvy customers are also the most motivated to bypass external dedicated suppliers.

The division of work in the AWS partnership is especially worth noting. In simple terms, a model first needs to understand the user's input, then gradually generate the response. The first part is more like "reading the context," and the second part is more like "quickly producing the next word." Trainium emphasizes the first part more, while Cerebras emphasizes low-latency output. This suggests Cerebras may not be replacing the entire GPU or cloud-chip path, but seeking to become a dedicated acceleration layer in the low-latency output segment.

This positioning is narrower, but also clearer. Cerebras does not need to prove that every AI task cannot function without it; it only needs to prove that a group of high-value, high-experience, sustainably used inference tasks need it over the long term.

5.6 Chapter Summary: What You Are Buying Is the Low-Latency Inference Budget Pool

The first part is not about solving "how large is the AI compute market," but about identifying which budget pool Cerebras actually occupies. The conclusion is clear: it should not be understood as a representative of the entire AI compute market, but as a company competing for value in the layers of low-latency inference, dedicated capacity, and inference interfaces.

The later competition chapter will separately break down how Nvidia, cloud vendors' in-house chips, Groq, custom ASICs, compute clouds, and physical infrastructure can split this money. For now, keep one judgment: Cerebras' opportunity comes from low-latency inference becoming an independent budget, and its risk comes from this budget being jointly diverted by customer gateways, cloud channels, the default GPU ecosystem, and delivery costs.

Chapter 6: What Cerebras Actually Is: Not a Big Chip, but Low-Latency Inference Infrastructure

Understanding Cerebras cannot start with "how big is the chip." A better starting point is: why would customers pay for faster AI responses, and how does Cerebras turn that speed into a service that can be bought, delivered, and billed?

In one sentence, Cerebras' company positioning is this: starting from wafer-scale hardware, it is trying to turn low-latency inference into infrastructure that can be purchased, called, and distributed.

First, break the company into the simplest economic-machine table. What it sells is not "chip size," but inference capability that is faster, more stable, callable, and deliverable to customers.

Question to Ask	Cerebras' Answer	How Investors Should Read It
What exactly does it sell?	CS-3 systems, low-latency inference capacity, Inference API, model services, code products, and compute backend in partner channels	There are many product forms, but ultimately all must return to capacity, usage, revenue, gross margin, and cash
Who pays?	Model companies such as OpenAI, cloud platforms, enterprise developers, research and government customers, and application providers that need real-time AI workflows	OpenAI is the largest anchor, but platformization depends on whether non-OpenAI customers replicate
Why pay?	Customers want faster responses, shorter agent loops, more stable real-time interaction, and higher task completion efficiency	Speed must become productivity within workloads, not just remain in benchmarks
When is revenue recognized?	Only after systems or capacity reach a service-ready state and customer acceptance or the service period satisfies recognition conditions does it enter revenue	The future revenue contract pool is visibility, not cash already received
Where do costs appear first?	Chips, systems, data centers, power, cooling, networking, hosting, depreciation, operations, and engineering teams	This is not an asset-light software business; growth consumes capital first
How does customer demand become per-share value?	Low-latency demand -> available capacity -> revenue -> gross profit -> operating cash flow -> free cash flow after capital expenditures -> diluted per-share value	Only if this path closes do common shareholders truly benefit

Layer	What It Is	What Customers Buy	What to Watch Financially	Current Boundary
WSE-3 / CS-3	Wafer-scale chips and systems designed around them	Dedicated AI compute systems, private deployments, low-latency capacity	System revenue, hardware gross margin, inventory, PP&E, depreciation	Needs to prove delivery stability and hardware gross margin
Inference API	High-speed inference cloud service	Token calls, low latency, model access	Usage revenue, gross margin, usage volume, customer concentration	Real call volume and unit economics are not fully disclosed
AI Model Services	Model development and enterprise services	Custom models, fine-tuning, proprietary data access	Service revenue, labor cost, project reuse rate	Can easily become labor-intensive projects
Cerebras Code	Developer workflow gateway	Code generation, code-agent experience	Subscription/usage, retention, capacity constraints	The reason for sold out still needs validation
Partner distribution	AWS, Hugging Face, OpenRouter, Poe, Vercel, and other gateways	Lower access friction and developer distribution	Channel share, customer gateway ownership, pricing power	Expanded distribution does not equal profit ownership

6.1 Why AI Gets Stuck on Speed

When a user asks AI a question, the model usually does not output the entire answer all at once. It typically generates one token at a time. For each token generated, the system has to move model parameters, context, cache, and compute results among compute units, memory, and networks.

Therefore, inference is not simply "calculating fast." In many cases, what truly slows the experience is data movement and waiting chains. What users see is a pause; what happens behind the system is scheduling, reading and writing, transmission, computation, and then transmission again.

What Users See	What Happens Behind It	Why It Matters
AI replies feel delayed	Before the first token appears, the system has to schedule the model, read context, and prepare computation	Determines whether users feel it is "instant"
Answers are generated slowly in segments	The model has to continuously generate the next token	Determines the speed of long answers, code, and multi-step tasks
Multi-agent tasks are especially slow	A task is split into multiple calls, and every step has to wait	Every small delay is amplified across the full task chain
Voice assistants feel unnatural	Voice interaction cannot tolerate obvious pauses	Low latency directly affects whether the product is usable
Code assistants stall	Developers need real-time suggestions while writing code	Slow speed makes the tool regress from a "workflow" into an "occasional query"

Traditional GPU clusters are extremely strong, especially at processing a large number of requests together to achieve high aggregate throughput. But that does not mean every user, every interaction, and every task chain is fastest. A system can be like a large bus that carries many people at once, or like a private car that gets a single passenger there faster. Cerebras is trying to do part of the latter: provide faster responses for high-value tasks whose users are willing to pay for speed.

6.2 Why SRAM and DRAM Matter

SRAM and DRAM are a simple entry point for understanding Cerebras' technology path, but the boundary must be made clear first. Simply put, AI inference is not only a compute problem; it is also a problem of how far data is from the compute units. The farther away data is, the more waiting there is; the more waiting there is, the slower the answer users see.

DRAM or HBM is more like a larger external warehouse. It can hold a lot of data, but each time the data has to be moved out, sent to compute units, and then the result has to be sent back. SRAM is more like a small drawer next to the workstation. Its capacity is much smaller and cost is higher, but access speed is faster. Cerebras' path is to place a large amount of on-chip SRAM and communication structure on an entire wafer-scale chip, minimizing data movement across chips, machines, and networks.

This is why Cerebras emphasizes on-chip memory, on-chip communication, and low latency. It is not simply saying "the chip is larger, therefore better." It is saying that if a class of AI tasks repeatedly reads model weights, context, and intermediate results, keeping more data closer may reduce waiting time.

Variable	More Intuitive Understanding	Why Cerebras Emphasizes It	What Investors Need to Keep Verifying
SRAM	Small-capacity high-speed memory very close to compute units	Helps reduce waiting and data movement	Whether capacity limits affect large models and long context
DRAM / HBM	Larger-capacity external high-speed warehouse	GPU clusters rely on larger external memory and high-bandwidth interconnect	Whether external memory and network optimization will narrow the gap
On-chip communication	Data moves within the same wafer	Reduces cross-card and cross-machine communication cost	Whether the advantage declines once a task leaves the wafer
Low latency	Users see the first answer faster, and multi-step tasks feel smoother	Code agents, voice, and real-time search can perceive it more easily	Whether customers are willing to keep paying for this experience

There is also an easy misread here: SRAM is not a universal advantage. It is fast, but expensive and capacity-limited. For small-batch, highly interactive, low-latency tasks, Cerebras' design may be very valuable; for extremely large models, long context, batch throughput, and already highly optimized GPU workloads, the boundary of the advantage still needs to be validated by real benchmarks, customer usage, and gross margin.

6.3 How the Technology Stack Is Layered

Cerebras' technology cannot be summarized only as "the chip is big." A more accurate breakdown has three layers.

The first layer is WSE-3 and CS-3. WSE-3 is a wafer-scale processor, and CS-3 is the system built around it. Key parameters disclosed in the company's datasheets and white papers include: WSE-3 uses TSMC 5nm, about 4 trillion transistors, about 900,000 active cores, 44GB SRAM, and 21PB/s memory bandwidth; CS-3 offers up to 125 petaFLOPS sparse FP16, 16RU, up to 27kW, and about 54kW for a dual-system rack. These numbers are not for showmanship, but to explain why the company emphasizes on-chip memory, on-chip communication, and low latency.

The second layer is MemoryX, SwarmX, and weight streaming. They mainly serve the training path: separating model-weight storage, broadcasting, and gradient aggregation from the compute units, so CS-3 becomes more like a system dedicated to matrix computation. The training chain and inference chain cannot be mixed together. Training capability shows the company has deep system engineering, but the current investment main line is closer to OpenAI capacity, high-speed inference, and API commercialization.

The third layer is Inference API, AI Model Services, Cerebras Code, and partner interfaces. This layer packages the underlying hardware capability into products customers can call. What investors should watch is whether customers call it continuously, whether usage grows, whether pricing is stable, and whether gross margin remains.

Technology Layer	What Problem It Solves	Commercial Meaning	What Still Needs Validation
WSE-3 / CS-3	Minimizes data movement across chips and machines	Supports low latency and system differentiation	Mass-production cost, delivery stability, hardware gross margin
MemoryX / SwarmX / weight streaming	Weights and communication complexity in training	Preserves upside in training systems	Whether it forms replicable training revenue
Inference API / Model Services	Turns system capability into a service customers can call	Closer to usage revenue and customer workflows	Usage, retention, channel share, unit economics

The key point is this: Cerebras has not eliminated complexity; it has moved complexity. It reduces some communication complexity in GPU clusters, while adding complexity in wafer-scale manufacturing, liquid cooling, dedicated software, data centers, and delivery economics.

6.4 Technology Advantages and Boundaries

Cerebras' strength is not traditional cluster throughput, but the speed felt by a single user, including how many tokens a single user gets per second, how long the first answer takes, and whether continuous interaction is smooth. It places a large amount of compute and on-chip memory inside the wafer, reducing the multi-card communication, network coordination, and data movement problems common in GPU clusters. For small-batch, highly interactive, low-latency inference tasks, this architecture can be very valuable.

Technical Advantage	Why It Matters	Still Needs Validation
wafer-scale architecture	Reduces multi-card, multi-machine, multi-network communication	The boundary of the advantage depends on model and workload
On-chip communication and on-chip memory	Helpful for low-latency interaction	Large models and long context may need more external resources
High tokens/sec/user	The user-perceived speed is better	Whether it can become payment and retention
OpenAI-compatible API	Lowers the developer trial threshold	Enterprise-level usage and revenue are not fully disclosed
Large-customer capacity contract	The technology is not just a laboratory result	Delivery and gross margin still have to be proven

The technology risk is not "there are no highlights," but that the market extrapolates those highlights too broadly. Cerebras may be very strong in a subset of low-latency inference tasks, but that does not mean it can cover all training, all inference, all model sizes, and all customer environments.

Pressure	Meaning	What to Watch
Limited SRAM capacity	Large models, long context, and KV cache may weaken the advantage	Supported model range, context length, real-task performance
off-wafer I/O limits	Once tasks frequently leave the wafer, the speed advantage may decline	Latency curves under different models and batch sizes
Multi-wafer pipeline	Extremely large models require cross-system collaboration, and complexity reappears	Scaling efficiency for large models
GPU / TPU / ASIC rapid optimization	Competitors will also optimize low-latency inference	Nvidia / TPU / Trainium inference updates
API usage scale not disclosed	Speed experience does not equal enterprise-level revenue	API revenue, usage, retention
Long-term utilization unknown	If dedicated capacity is underutilized, unit economics will suffer	Revenue per MW, gross margin, customer usage stability

A more rigorous technology conclusion is not "Cerebras is faster than GPUs," but this: Cerebras has a clear architectural advantage in certain low-latency inference scenarios, but the boundary of that advantage depends on model size, context length, batch size, concurrency, I/O, customer migration cost, and actual utilization.

---

Chapter 7: Demand Pool and Value Pool: Why AI Inference Growth Does Not Equal Company Cash Flow

Growth in AI inference demand is a real backdrop, but it is not enough for the stock. Investors need to separate three things: who gains value from low latency, who pays for low latency, and who ultimately keeps the profit.

Workload	Why Speed Matters	Representative Customers / Scenarios	How Investors Should Read It
Large-scale low-latency capacity	Large-model customers need to provide low-latency capability to their own end users	OpenAI	The strongest revenue visibility and the greatest delivery pressure exist at the same time
Cloud distribution	Developers need lower access friction	AWS / Bedrock, Meta Llama API, Hugging Face, OpenRouter	Proves ecosystem access, but the customer gateway may be captured by channels
Code agents	Slowness interrupts developers' continuous workflow	Cognition / Windsurf, Cerebras Code, Codex-Spark	Speed affects not just waiting, but whether developers are willing to hand tasks to agents
Enterprise search	The more documents and citations there are, the more obvious waiting becomes	Notion, AlphaSense	Low latency may increase follow-up frequency and depth of use
Voice/video agents	Human conversation is extremely sensitive to pauses	Tavus, OpenCall, Sei, Norby, Delphi	The experience is real, but financial evidence is weaker than contract evidence
Life sciences and research	The number of iterations affects experimental progress	GSK, Mayo, AstraZeneca, Argonne, nference	Proves specialized workloads, but the commercial cycle is longer
National labs / sovereign AI	Owned capacity, privacy, and research control matter	DOE, Sandia, LLNL, PSC, EPCC, G42	Proves system deployment and budget clues, but procurement scale and renewals must be watched

Customer cases should be treated as workload proof, not revenue proof. They can show "why someone may need Cerebras," but they cannot directly show "how much revenue, gross margin, and cash flow these customers contribute."

Who Ultimately Takes the Profit

Low-latency inference may improve user experience, OpenAI may gain stronger product capability from it, AWS may gain a more complete cloud service, and developers may gain a faster API. What common shareholders truly care about is how much of that value Cerebras can retain through contracts, pricing, gross margin, and cash flow.

Where Value Occurs	Who Benefits Most Directly	What Cerebras Must Control to Retain Value
Low-latency experience	OpenAI, developers, end users	It cannot only be a one-time hardware supplier; it must make speed a continuously billed service
Capacity goes online	OpenAI and other large customers	Contracts must clearly bind capacity responsibility, payment timing, and service acceptance
Cloud distribution	AWS, Meta Llama API, and other platforms	It cannot become only a hidden backend; it must retain pricing, usage, and renewal negotiation rights
Model services	Enterprise customers, developer platforms	API usage must keep growing, rather than relying only on large-customer contracts
Hardware systems	Data centers, hosting, power, equipment supply chain	It must control pass-through and reimbursed costs, depreciation, operations, and utilization risk
Equity binding	OpenAI, employees, early investors	Common shareholders need to look at diluted per-share cash flow, not only enterprise scale

Therefore, technology leadership cannot be directly equated with a good stock. Cerebras can be clearly faster in certain tasks, but if customers push down the speed premium, cloud platforms take the billing relationship, or OpenAI warrants and equity compensation consume the per-share allocation, the value received by common shareholders will be diluted.

---

Chapter 8: The OpenAI Contract: Revenue Visibility, Delivery Responsibility, and Per-Share Dilution Appear Together

The key fact in this Cerebras report is not "it has a large customer," but that the OpenAI contract has pushed the company into a new position: Cerebras is no longer only proving that its chip is fast. It must prove that it can deliver 750MW of inference capacity on schedule and turn that delivery into revenue, gross profit, cash, and diluted per-share value.

This is also where misreadings are easy. The larger the OpenAI contract, the stronger the revenue visibility; but the same contract also amplifies delivery responsibility, customer concentration, capital expenditures, loan constraints, and warrant dilution. It is not a pure positive, nor a pure risk, but a master contract that ties together all of Cerebras' core questions for the next several years.

8.1 This Is Not an Ordinary Order

According to the OpenAI MRA, loan documents, warrant documents, and 424B4, the relationship between OpenAI and Cerebras has already gone beyond ordinary procurement. It simultaneously includes capacity commitments, engineering governance, fee mechanisms, customer financing, exclusivity arrangements, and equity linkage.

Contract Component	Disclosed Fact	Benefit to Cerebras	What Investors Cannot Ignore
Capacity commitment	OpenAI initially committed to purchase 750MW of inference capacity: 250MW in 2026, 500MW cumulative in 2027, and 750MW cumulative in 2028	Provides a strong anchor for the future revenue curve	Contracted capacity does not equal available capacity already online
Service nature	The contract points to end-to-end high-speed inference services for OpenAI model inference	Proves Cerebras is not only selling one-time hardware	Services must reach an available state, acceptance, and continuous operation before revenue is recognized
Payments and fees	Fees include system fees, professional service fees, and certain pass-through/reimbursed costs; payment obligations have non-cancelable, non-refundable features, but several details are redacted	Improves revenue visibility	Total revenue quality depends on pass-through costs, audit rights, gross margin, and collections
Joint governance	The parties set up governance committees covering data-center procurement, forecasting, capacity planning, engineering collaboration, fees, and operating decisions	Shows OpenAI is deeply involved in production scheduling and engineering execution	This also means delivery timing will be monitored more strictly
Customer loan	OpenAI provides an approximately $1.0045 billion secured promissory note, 6% interest rate, due 2032, with funds going into a lockbox account	Supports capacity buildout and reduces short-term financing pressure	The loan is not revenue, use of funds is restricted, and it must later be repaid or offset through services/assets
Warrants	OpenAI can receive up to 33.445M Class N shares at an exercise price of $0.00001, with portions tied to the loan, valuation, capacity delivery, and additional capacity	Deepens alignment between customer and company	The more successful the demand validation, the more potential dilution must enter per-share value calculations
Exclusivity arrangements	The contract contains exclusivity mechanisms related to Named Competitors, but the list, term, and some conduct are redacted in public filings	May increase cooperation stickiness	It cannot be written as a definite ban on a specific competitor, and it may also limit some commercial space

The core of this contract is not "how much OpenAI bought," but "OpenAI has turned Cerebras into an important inference-capacity buildout partner." That sentence is more valuable, and also more dangerous. More valuable because a top-tier customer is willing to entrust key capacity to it; more dangerous because Cerebras' execution is no longer limited to laboratories and demos, but must undergo continuous testing by data centers, power, cooling, networking, system reliability, service acceptance, and financial statements.

8.2 750MW Is Not a Number, but a Delivery Timeline

The meaning of 750MW is not that it sounds large, but that it turns Cerebras' revenue recognition schedule, capital investment schedule, and customer trust schedule into a timeline.

Time Point	Contract Capacity Schedule	What Must Happen Operationally	What to Watch in the Financial Reports
2026	250MW	Data centers, power supply, cooling, networking, and CS systems must gradually enter a service-ready state	One-year recognizability within the future revenue contract pool, revenue growth, receivables, contract assets/liabilities, capital expenditures
2027	500MW cumulative	If the first phase of delivery goes well, customer usage and subsequent expansion should become clearer	Whether gross margin is stable, whether operating cash flow improves, and whether data-center costs spill over
2028	750MW cumulative	Large-scale capacity delivery enters fuller validation	Whether revenue is recognized on schedule, whether depreciation and operating costs are bearable, and whether customer concentration declines
After 2029	Additional capacity options exist, but specific MW and some terms are redacted	Additional capacity becomes more valuable only after early delivery succeeds	Redacted terms cannot be placed fully into the base case in advance

This timeline reminds investors that the next few years for Cerebras are not about "waiting for the AI megatrend to keep growing," but about whether capacity nodes one by one can become available services. If any node is delayed, the impact is not only a news item; revenue recognition, gross margin, cash flow, loan repayment, warrant vesting, and stock-price expectations all come under pressure together.

8.3 How to Read the Billion-Dollar Future Revenue Contract Pool

Multiple materials citing the S-1 show that Cerebras' future revenue contract pool is about $24.6 billion, with a significant portion related to the OpenAI MRA; a common conversion framework is about 15% recognized in 2026-2027, about 43% in 2028-2029, and the remainder recognized later. This framework can be used to judge revenue visibility, but it cannot be treated directly as profit, cash, or per-share value.

A more accurate reading is to treat it as a service commitment that still must pass six gates.

Link That Must Be Passed	What It Shows	If It Is Not Passed
The contract remains valid	Customer commitments, payment obligations, and governance mechanisms remain in place	Revenue visibility declines, and the market will reassess contract quality
Capacity reaches a service-ready state	Data centers, systems, power, cooling, networking, and security can all support service	The contract exists, but revenue recognition shifts later
Customer acceptance and continuous usage	Service performance meets customer requirements and customers make real calls	Capacity may sit idle, or service credits and renegotiation may be triggered
Revenue is recognized under the rules	The future revenue contract pool enters the income statement	The growth curve is below market expectations
Gross profit is retained	There is still profit after pass-through costs, power, hosting, depreciation, and operations	Revenue grows but gross margin is low, weakening the valuation logic
Cash and per-share value come back	After collections, loans, capital expenditures, warrants, and equity compensation, common shareholders still benefit	The company becomes larger, but per-share value is diluted or offset by cash consumption

Therefore, the roughly $24.6 billion future revenue contract pool is a very strong starting point, not the finish line. It can make the market study Cerebras seriously, but it cannot allow investors to skip delivery, gross margin, cash, and the equity denominator.

8.4 How This Contract Enters the Three Financial Statements

The OpenAI contract should be the main line of the full report because it changes the income statement, balance sheet, and cash flow statement at the same time. It is not just a customer news item that affects revenue.

Financial Position	How the OpenAI Contract Enters	Most Positive Interpretation	Most Dangerous Interpretation
Income statement revenue	After 750MW capacity reaches a service-ready state, service revenue is recognized under the rules	Large-customer revenue begins to materialize	Slow delivery, slow acceptance, or shortened service periods push revenue later
Income statement gross profit	System fees, professional services, pass-through costs, power, hosting, depreciation, and operations together determine gross profit	Service scaling effects emerge	Pass-through and reimbursed costs are high, creating large revenue but weak gross profit
Balance sheet assets	Data centers, hardware, systems, prepayments, contract assets, fixed assets, and possible right-of-use assets increase	The company forms reusable inference infrastructure	Assets become too heavy, utilization is insufficient, and depreciation pressure rises
Balance sheet liabilities	OpenAI loan, potential contract liabilities, leases, or other delivery obligations increase	The customer helps the company build capability	Debt and restrictive arrangements reduce financial flexibility
Cash flow statement	Loans and prepayments may improve cash first, while construction, operations, and capital expenditures later consume cash	Funds support capacity expansion	Cash improvement comes from financing rather than operations, while operating cash flow remains unstable
Per-share value	OpenAI warrants, employee equity compensation, options, and lock-ups jointly affect the denominator	Customer alignment improves long-term demand credibility	Warrant vesting and equity compensation reduce the value allocated to common shareholders

Therefore, looking only at revenue growth will misread Cerebras. The question is whether the revenue growth brought by the same OpenAI contract can outpace asset expansion, liability constraints, cash consumption, and equity dilution.

8.5 What the Loan and Lockbox Account Show

OpenAI provides an approximately $1.0045 billion loan, with a 6% interest rate and maturity in 2032; the related funds enter a lockbox account and are used to accelerate services, software development, hardware manufacturing, and related buildout. This arrangement gives Cerebras more funding support for capacity construction and also shows OpenAI is deeply involved in the delivery schedule.

But this money cannot be treated as operating cash flow. It is more like restricted financing provided by the customer to bring future services online faster. For investors, the key is not "the company has more money," but how this money is used, when it will be repaid, whether it is offset through service delivery or asset arrangements, and whether it makes Cerebras' operating cash flow look stronger than it actually is.

Reading	Correct Understanding
The loan improves the safety cushion	Yes, but it has an interest rate, maturity date, lockbox account, and repayment arrangement
The loan equals customer prepayment	It cannot be written this way. The loan is a financing arrangement, not revenue
The loan proves OpenAI trusts Cerebras	It can be treated as a strong signal, but delivery still has to occur on schedule
The loan will improve cash flow	It may improve short-term funding status, but does not mean operating cash flow has improved

The key observation point in future quarterly statements is whether cash increases come from customer financing, IPO proceeds, prepayments, or operating cash generated by service revenue itself.

8.6 Warrants Tie Demand Validation and Per-Share Dilution Together

The OpenAI warrants are the easiest part of this contract to overlook. They correspond to up to 33.445M Class N shares at a very low exercise price; part has already vested because the loan was funded, and other portions are tied to market value, payments, capacity delivery, and additional capacity.

This means a somewhat non-intuitive but important fact: the more successful the OpenAI-related business is, the more likely the related warrants are to enter the equity denominator pressure. In other words, demand validation and common-share dilution may strengthen at the same time.

OpenAI Warrant-Related Situation	What It Says About the Business	What It Says for Common Shareholders
Loan triggers partial vesting	OpenAI has already put funding into the delivery system	Denominator pressure begins to appear
Market value or payment triggers partial vesting	Market and customer payments enter a stronger validation stage	Per-share value needs to be diluted again
Capacity delivery triggers partial vesting	The more fully the OpenAI contract is executed	Enterprise value may rise, but per-share value depends on the diluted result
Additional capacity triggers more vesting	The contract ceiling becomes larger	Investors cannot look only at the revenue ceiling; they also need to look at the warrant denominator

This is not to say the OpenAI warrants are necessarily bad. They may be part of the cost of enabling the large contract and customer alignment. The issue is that investors must separate them from the "good contract news" and put them into a per-share value stress test.

8.7 Exclusivity Arrangements Need to Be Written Carefully

The public version of the contract discloses exclusivity mechanisms related to Named Competitors, but the specific list, term, and some restricted conduct are redacted. This fact can show the relationship between the parties is deeper, and may also show that commercial expansion has boundaries; but it cannot be written as "OpenAI clearly cannot use a specific competitor," and it cannot be written as "Cerebras clearly cannot serve a specific customer."

The correct reading is this: the exclusivity arrangement increases cooperation stickiness and also increases information opacity. It makes the OpenAI contract look more like a deep strategic partnership than ordinary procurement; at the same time, it requires investors to remain disciplined when judging customer replication, because the public filings do not tell us which customers or competitive paths are restricted.

8.8 OpenAI Is a Strong Endorsement, but Cannot Become the Whole Story

Cerebras' strongest evidence today comes from OpenAI; one of its largest risks also comes from OpenAI. Strong endorsement can support high expectations, but if future revenue remains highly dependent on a small number of customers such as OpenAI, G42, and MBZUAI, the company looks more like a special-project supplier than a broad platform.

Future judgment should focus on four data points instead of repeatedly quoting the total contract amount.

Data to Watch	Why It Matters	Positive Signal	Negative Signal
Rolling table for future revenue contract pool	Judges whether the contract pool enters revenue on schedule	The portion recognizable within one year rises, and the recognition schedule is clear	The contract pool shifts later, explanations are vague, or contracts are renegotiated
Capacity go-live progress	Judges whether 250MW / 500MW / 750MW lands	Site, power, cooling, system, and acceptance timing are transparent	There are only contract numbers, with no capacity online
Non-OpenAI revenue	Judges whether the platform is replicable	API, AWS, enterprise, and developer revenue become visible	OpenAI still explains the vast majority of growth
Diluted per-share metrics	Judges whether common shareholders truly benefit	Revenue, gross profit, and operating cash flow grow faster than the equity denominator	Enterprise scale expands, but per-share revenue and per-share cash are diluted

In one sentence, the OpenAI contract has put Cerebras on the main stage, but the brighter the stage lights, the stricter the financial-statement validation. Future investment judgment cannot stop at "the contract is large." It must watch whether this contract can pass through delivery, gross margin, cash, and dilution, and still make common-share per-share value larger in the end.

---

Chapter 9: Product-Line Economics: How Systems, Inference Interfaces, Model Services, Code Products, and Partner Channels Make Money

Cerebras does not have only one hardware revenue category. It has at least five commercial entry points. Each entry point enters the three financial statements differently, and investors cannot blend them into one "AI compute revenue" bucket.

Product Entry Point	Who the Customer Is	How Revenue Is Generated	Where Costs Appear First	What It Can Prove	What It Cannot Prove
CS-3 / WSE systems	Large model companies, national labs, enterprise/government compute customers	System sales, dedicated capacity, project deployment	Chips, systems, data centers, depreciation, delivery services	Cerebras has a deliverable hardware and system path	It does not automatically prove high gross margin and high utilization
Inference API	Developers, enterprise applications, model service platforms	Token / API calls or service fees	Inference capacity, operations, power, model-service costs	The company can move from hardware to callable services	It does not prove usage, retention, and unit economics
AI Model Services	Enterprise, government, and research customers	Model training/fine-tuning/deployment services	Engineering teams, system capacity, delivery cycles	It has the potential to move up into solutions	It does not prove scalability and repeatability
Cerebras Code	Developers and code-agent users	Subscriptions, API calls, or channelized products	Models, inference capacity, product R&D	Low latency has perceptible value in agent workflows	Sold out does not equal a revenue breakout
Partner channels	AWS, OpenRouter, Hugging Face, Vercel, etc.	Distribution, API calls, platform integrations	Revenue share, channel relationships, customer support	Access friction declines and customer reach expands	The customer gateway and gross margin may be taken by channels

More product entry points do not equal more profit. The questions are who owns the customer relationship, who sets pricing, who bears capacity cost, and who ultimately recognizes gross profit.

How Product Entry Points Enter the Three Financial Statements

Product Entry Point	What the Income Statement Will Show	What the Balance Sheet Will Show	What the Cash Flow Statement Will Show	What Most Needs Validation
CS-3 systems	Hardware/system revenue and gross margin	Inventory, receivables, PP&E, contract assets/liabilities	Production, delivery, installation, and capital expenditures consume cash	Whether hardware gross margin is stable and delivery is on schedule
Inference API	Usage revenue, subscription, or enterprise API revenue	Receivables, contract liabilities, possible capacity assets	Cash outlays for data centers, power, hosting, and operations	Whether API usage is sustained and per-token gross margin works
AI Model Services	Service revenue and project delivery costs	Contract assets, deferred revenue, receivables	Labor, engineering delivery, and customer support spending	Whether projects are reusable and can become high-margin services
Cerebras Code	Subscription or usage revenue	Deferred revenue and customer contracts	Capacity costs, product support, and development investment	Whether sold out means strong demand, capacity limitation, or a small-scale pilot
Partner distribution	Channel revenue or net revenue	partner settlement, receivables, and contract arrangements	Revenue share, platform fees, and capacity costs	Who owns the customer gateway, who owns pricing power, and who retains gross margin

The conclusion here is not "more gateways are better." More gateways lower adoption friction, but profit ownership needs to be unpacked more carefully. Platforms such as AWS, Hugging Face, OpenRouter, Poe, Notion, and AlphaSense can help Cerebras enter more workloads, and they can also keep the customer relationship and some profit for themselves.

---

Chapter 10: What the Three Financial Statements Show Together: Revenue, Assets, Cash, and Per-Share Value

Cerebras' financial statements cannot be read like an ordinary chip stock, and cannot be read like an asset-light software stock. It is tying together wafer-scale hardware, system delivery, data-center capacity, and inference services. If this model works, it could become a high-value AI inference infrastructure company; but before it works, the three statements will show pressure first.

The signal currently given by the three statements is clear: demand visibility is very strong, and revenue has already grown, but the financial statements have not yet proven that Cerebras can steadily turn large contracts into high-quality gross margin, operating cash flow, and diluted per-share value. It has passed the "are there customers" stage and is entering the "how do customer commitments become shareholder cash" stage.

10.1 How This Company Turns Customer Demand Into Per-Share Value

Cerebras' cash path is not "sell a chip and it is over." The more accurate path is: customers first commit to purchase inference capacity; the company then builds systems and data-center capability; after services reach an available state, revenue is recognized; only after revenue deducts hardware, hosting, power, depreciation, and operating costs does gross profit form; gross profit then passes through operating expenses, capital expenditures, and equity dilution, and only then may become per-share value.

Business Action	What the Income Statement Will Show	What the Balance Sheet Will Show	What the Cash Flow Statement Will Show	How Common Shareholders Should Read It
Sign long-term capacity contracts	Current-period revenue may not immediately increase, while future revenue visibility improves	The future revenue contract pool increases, possibly accompanied by changes in contract assets, contract liabilities, or receivables	Does not necessarily bring operating cash flow immediately	Contracts are the starting point of value, not the cash endpoint
Build or procure AI capacity	R&D, operations, depreciation, and delivery costs gradually enter expenses or costs	Fixed assets, leases, prepayments, inventory, supplier commitments, or restricted cash increase	Capital expenditures and construction cash flow out first	The faster the growth, the greater the possible short-term cash pressure
Capacity reaches a service-ready state	Revenue begins to be recognized, and gross margin is tested	The future revenue contract pool gradually turns into revenue, while receivables and contract accounts change	Collections, operating costs, and working capital together determine operating cash flow	This is the first gate from contract to business
Customer loan or financing support	Interest and accounting treatment may affect the income statement	Liabilities increase, and lockbox accounts or restricted funds may appear	Cash inflow, but not operating cash	Financing cash cannot be treated as business self-funding
Issue equity to customers or employees	Equity compensation and warrants affect expenses or dilution metrics	The equity denominator expands and the equity structure becomes more complex	Does not directly generate operating cash	Company value growth must outpace denominator expansion

This path explains why Cerebras cannot be judged only by revenue, and cannot be judged only by the future revenue contract pool. The question is whether revenue growth can outpace growth in assets, liabilities, capital expenditures, and the equity denominator.

10.2 Income Statement: Revenue Has Expanded, but Gross Profit and Operating Profit Have Not Expanded in Sync

The first signal from the income statement is positive: Cerebras is no longer a concept company with no revenue. 2025 revenue was about $510 million, up about 76% from about $290.3 million in 2024. This shows demand is real and the company can turn part of customer demand into revenue.

But the second signal from the income statement is more important: gross profit grew slower than revenue, gross margin declined, and operating loss widened. In other words, revenue is expanding, but costs and expenses are also expanding, and scale effects have not yet appeared consistently.

Item	2024	2025	Change	Reading
Revenue	Approx. $290.3 million	Approx. $510.0 million	+approx. 76%	Demand and delivery scale have expanded significantly
Gross profit	Approx. $122.7 million	Approx. $199.1 million	+approx. 62%	Gross profit growth is slower than revenue growth
Gross margin	Approx. 42.3%	Approx. 39.0%	Down approx. 3.3 percentage points	Revenue growth has not yet brought gross-margin improvement
Operating loss	Approx. -$101.4 million	Approx. -$145.9 million	Loss widened	The company is still using R&D, sales, delivery, and infrastructure investment to exchange for scale
GAAP net income	Approx. -$481.6 million	Approx. +$237.8 million	Superficially turned positive	Mainly affected by a one-time accounting gain, and cannot be treated as mature operating profitability

The easy misread here is GAAP net income turning positive. 2025 net income appears positive on the surface, but public materials show it included an approximately $363.3 million one-time accounting gain. This gain can affect accounting profit, but it cannot replace operating profit, gross margin, and cash flow. For Cerebras, the income statement line to watch is not "whether net income turned positive," but whether revenue, gross profit, operating loss, and future depreciation pressure can improve together.

10.3 Balance Sheet: Moving From Chip Company to Infrastructure Company Makes Assets and Obligations Heavier

The OpenAI 750MW capacity commitment changes the nature of Cerebras' balance sheet. It is no longer only about inventory, receivables, and cash. In the future, investors must also watch fixed assets, data-center commitments, leases, prepayments, contract assets, contract liabilities, restricted cash, customer loans, and the equity denominator.

Balance Sheet Position	Why It Becomes Important	Positive Signal	Negative Signal
Receivables and contract assets	Whether collections arrive on time after revenue recognition	Revenue grows while receivables turnover remains controlled	Revenue growth comes with slower collections
Contract liabilities or customer prepayments	Whether customers pay upfront to support delivery	Prepayments match subsequent revenue recognition	Cash flow relies on one-time prepayments and is not sustainable later
Fixed assets and right-of-use assets	Whether data-center, hardware, and system construction becomes heavier	Asset turnover improves, and each dollar of assets generates more revenue	Assets grow faster than revenue, and depreciation pressure rises
Inventory and supplier commitments	Whether hardware delivery requires advance stocking	Inventory matches the delivery schedule	Inventory piles up or supplier commitments lock up cash
OpenAI loan and lockbox	How customer financing supports capacity buildout	Funds drive service-ready capacity online	Use of funds is restricted, and repayment arrangements constrain financial flexibility
Warrants, options, and equity compensation	Whether per-share value will be diluted	Operating delivery outpaces equity expansion	Company value growth is consumed by denominator expansion

The balance sheet should answer the most important question: how heavy must Cerebras' assets and obligations become to deliver the future revenue contract pool? If the future revenue contract pool is large, but each dollar of revenue requires more fixed assets, more leases, more prepayments, and more equity compensation to support it, common-shareholder returns will be compressed.

10.4 Cash Flow Statement: Operating Cash Flow Has Not Proven Stable Self-Funding

The cash flow statement currently gives a more cautious signal than the income statement. 2025 operating cash flow was about -$10 million, while 2024 operating cash flow was about +$452 million. The higher 2024 operating cash flow was affected by large customer prepayments and cannot be treated directly as stable operating cash generation.

2025 major capital expenditures were about $382.7 million. On a simple basis:

This does not mean Cerebras has no value; it means the company is still in a buildout phase. The real question is not "whether cash flow currently looks attractive," but whether operating cash flow can improve after future revenue recognition, whether capital intensity can decline, and whether diluted free cash flow per share can emerge.

Cash Flow Item	What It Shows Now	What to Watch Next
Operating cash flow	2025 has not yet proven stable self-funding	Whether it improves excluding prepayments, and whether receivables and contract assets are controlled
Capital expenditures	Expansion spending in 2025 is significant	Whether capital expenditures are a buildout peak or a long-term rolling reinvestment need
Customer loan	Provides delivery funding, but is not operating cash	Use of funds, lockbox restrictions, repayment method, and interest burden
IPO proceeds	Strengthen cash and delivery capability	Cash burn rate, RSU taxes, capital expenditures, and operating losses
Free cash flow	Still under pressure	Whether revenue, gross profit, and operating cash flow can outpace expansion spending

10.5 The Gaps Between Growth, Profit, and Cash

What Cerebras most needs to watch now is not a single metric, but the distance between groups of metrics. If these distances narrow, the company moves from project-style delivery toward platform-style capacity services; if these distances widen, the market will reprice it.

Gap	Current Direction	Possible Reason	Next Evidence	Investment Meaning
Revenue growth vs gross profit growth	Revenue approx. +76%, gross profit approx. +62%	Cost structure, project mix, pass-through and reimbursed costs, delivery costs	Whether gross margin can stabilize or rebound	Judge revenue quality, not just revenue scale
Revenue growth vs operating loss	Revenue grows, but operating loss widens	R&D, sales, delivery, and infrastructure investment come first	Whether operating expense ratio declines	Judge whether scale effects appear
GAAP net income vs operating income	GAAP turned positive, but operating loss remains	One-time accounting gain affects net income	Profit excluding non-recurring items	Avoid misreading accounting gains as operating maturity
Operating cash flow vs free cash flow	Operating cash flow is near zero, and capital expenditures are large	Expansion buildout consumes cash	Whether capital expenditures / revenue declines	Judge whether the business will keep relying on external financing
Future revenue contract pool vs revenue	The contract pool is large, while revenue is still in early recognition	Delivery, acceptance, and service periods determine recognition timing	One-year recognizable portion, rolling table	Judge whether contracts truly turn into reported revenue
Capacity commitments vs capacity online	OpenAI 750MW is large, while disclosed site samples remain limited	Data centers, power, cooling, and system installation are hard constraints	MW disclosed by site, service-ready state reached, acceptance	Judge whether contracted capacity becomes sellable services
Enterprise value vs equity denominator	Market pricing is high, and the dilution bucket is also large	IPO, OpenAI warrants, options, RSUs, plan reserves	Diluted shares, revenue per share, cash flow per share	Judge whether common shareholders truly benefit

This group of gaps is not meant to deny Cerebras, but to identify its real exam questions. Revenue growth is real; but if gross margin does not rebound, operating cash flow does not improve, capital expenditures continue consuming cash, and the equity denominator keeps expanding, revenue growth does not necessarily become a good investment.

10.6 How the Future Revenue Contract Pool, Capacity, and Capital Expenditures Reconcile

For Cerebras, the key is not one quarter's revenue, but whether the future revenue contract pool, capacity, assets, and cash can line up.

Link	Currently Known	What Still Needs Validation	Why It Matters
Future revenue contract pool	Approx. $24.6 billion long-term revenue visibility; common disclosure framework shows about 15% conversion in 2026-2027 and about 43% conversion in 2028-2029	Whether contracts are executed on schedule, whether the portion recognizable within one year improves, whether renegotiation or delays exist	Determines whether revenue visibility can enter the statements
Capacity	OpenAI initial capacity: 250MW in 2026, 500MW cumulative in 2027, 750MW cumulative in 2028	Power, cooling, networking, service-ready state, and acceptance progress at each site	Determines whether the future revenue contract pool can become revenue
Capital expenditures	2025 major capital expenditures were about $382.7 million	Whether this is a buildout peak or a sustained high-intensity requirement in future years	Determines cash flow pressure
Fixed assets and depreciation	Public materials are still insufficient to stably break out the future depreciation curve	Asset turnover, depreciation lives, asset impairment risk	Determines whether gross margin and return on capital can hold
Operating cash flow	About -$10 million in 2025; about +$452 million in 2024 but affected by customer prepayments	Whether it improves excluding prepayments, and whether receivables and contract assets are controlled	Determines whether the business can self-fund
Equity denominator	OpenAI warrants, options, RSUs, PSUs, and plan reserves can all affect per-share value	Whether diluted share count and per-share cash flow improve in sync	Determines whether common shareholders become wealthier after the company grows

This reconciliation chain needs continuous updating. The larger the future revenue contract pool, the more investors must watch whether assets become heavier; the heavier the assets, the more they must watch depreciation and operating cash flow; the weaker the cash flow, the more they must watch whether the company continues to rely on financing and dilution.

10.7 A Real-World Sample of Capacity Cost: Digi Power X 40MW

Cerebras' 40MW Alabama data-center contract with Digi Power X is an important sample for understanding the cost side. It is not Cerebras' revenue; it is part of the infrastructure resources Cerebras has locked in to deliver compute capacity.

Item	Fact	Financial Meaning
Capacity	40MW IT load	A small sample of the OpenAI 750MW target
Phasing	First phase 15MW, second phase incremental 25MW	Helps observe go-live timing
Term	Initial 10 years	Shows cost commitments are long term
Initial contract amount	Approx. $1.1 billion	A hosting and infrastructure cost anchor, not Cerebras revenue
First phase service-ready target	2026-12-15	Related to the 2026 capacity delivery schedule
Full deployment target	End of 2027 Q1	Related to 2027 delivery capability
Cost characteristics	Payment is required under the contract whether fully used or not, and Cerebras pays based on contracted IT load	If utilization is insufficient, cost pressure arrives first

40MW is only about 5.3% of OpenAI's 750MW. Even looking only at the 2026 target of 250MW, 40MW is only part of it. The next thing that most needs to be filled in is a site-by-site capacity list: how many MW each site has, when it goes online, who bears the cost, and what status allows revenue recognition.

10.8 Financial Signals That Are Easy to Overlook

Cerebras' risks may not appear first in news headlines, but more likely in small changes across the three statements.

Signal Easy to Overlook	Where It Appears	Why It Matters	Common Misread	What to Watch Next
Receivables and contract assets grow faster than revenue	Balance sheet	Shows revenue recognition and collections may not be synchronized	Only looking at revenue growth and not collection quality	Receivables turnover, changes in contract assets, bad debt or discounts
Contract liabilities decline but revenue does not keep up	Balance sheet and income statement	Shows that after prepayments are consumed, new cash has not kept up	Treating past prepayments as ongoing operating cash	Contract liability roll-forward
Fixed assets grow faster than revenue	Balance sheet	Shows expansion may be heavier than revenue release	Treating all capital expenditures as future growth	Revenue / fixed assets, depreciation pressure
Operating cash flow relies on financing or prepayments	Cash flow statement	Shows the business's own cash generation is still weak	Treating cash increase as operating improvement	Separate operating cash flow from financing cash flow
Gross margin is compressed by pass-through costs and depreciation	Income statement	Shows large revenue is not necessarily high quality	Treating the entire future revenue contract pool as high gross margin	Gross margin, service costs, depreciation, and hosting expenses
OpenAI warrants vest gradually	Equity footnotes	Shows demand validation and dilution may appear together	Looking only at contract success and not at the denominator	Diluted share count and revenue per share
Equity compensation and plan reserves expand	Equity footnotes	Affects per-share value	Looking only at enterprise value and not per-share value	SBC, RSUs, options, plan reserves

10.9 Which Numbers Need to Change for It to Move From Project Supplier to Platform

If Cerebras is only a project supplier, the three statements will show: a few large customers, large contracts, staged revenue, heavy capital expenditures, volatile cash flow, warrants, and financing dilution.

If it moves toward a platform, the three statements should gradually show: more customers, a continuously rolling future revenue contract pool, predictable delivery cycles, stable gross margin, positive operating cash flow, declining capital intensity, and improving diluted cash flow per share.

Signal	Project Supplier Characteristics	Platform Characteristics	Next Validation for Cerebras
Customer structure	One or two customers contribute most revenue	Customer count expands and single-customer concentration declines	Whether customers beyond OpenAI contribute substantive revenue
Future revenue contract pool conversion	Large contracts are hard to deliver on schedule	Converted into revenue on schedule	Whether approx. 15% conversion in 2026-2027 is realized
Gross margin	Each project's cost differs greatly, and gross margin fluctuates	Gross margin stabilizes after servicization and scaling	Whether gross margin can return to and exceed 40%
Capital intensity	Growth relies on continuous construction	Capital expenditure required per dollar of revenue declines	Revenue / fixed assets and revenue / capital expenditures curves
Operating cash flow	Relies on prepayments and has high quarterly volatility	Can turn positive even after excluding prepayments	Operating cash flow quality and changes in receivables
Equity denominator	Uses warrants, financing, and equity compensation to exchange for growth	Free cash flow per share begins to improve	Diluted share count and cash flow per share
Asset turnover	Capacity buildout slows returns	Capacity utilization improves and asset turnover accelerates	Conversion efficiency from MW to revenue

The conclusion of Part Six is simple: Cerebras' financial validation cannot stop at "revenue growth" and "a large contract pool." It must prove that revenue can cover expansion costs, assets can generate sufficient returns, operating cash flow can improve, capital intensity can decline, and the equity denominator will not eat away common shareholders' returns.

---

Chapter 11: IPO Supply, Lock-Up Releases, and Secondary-Market Absorption

Cerebras' post-listing trading cannot be judged only by the company story, and cannot be judged only by the first-day gain. The first stage of pricing for a new issue is often jointly determined by three things: how many shares can be bought, who holds low-cost shares, and how many shares will be released over the next several months.

Therefore, the question here is not "is the company worth researching," but a more specific question: for those buying near the first-day closing price of $311.07, what supply will they have to absorb in the future?

11.1 After Listing, Not Everyone Owns the Same Cost Basis

For the same company, different holders face entirely different cost bases and mental ledgers. IPO allocation investors bought at $185; some employees and early shareholders have economic costs far below the IPO price; first-day secondary-market buyers may buy above $300. This is not a moral judgment, but a trading-structure fact.

Holder Type	Typical Cost or Price	Main Motivation	What It Means for the Secondary Market
IPO allocation investors	$185.00	Receive first-day liquidity and a pricing discount	Already have meaningful first-day paper gains and may lock in profits
Directed share program participants	$185.00	Participate in the offering at the IPO price	Included in the 30.0m IPO shares, and usually not equivalent to long-term lock-up
Early-release shares from non-executive employees	Most costs may be below post-IPO prices	Liquidity, taxes, risk diversification	May create additional supply on the first and second trading days
Vested option holders	Weighted-average exercise price approx. $4.97	Exercise, pay taxes, monetize	There is a huge spread versus $311.07
Series G / tender reference price	Approx. $36.23	Reference for mid-to-late-stage investment and employee tender	Still about 8.6x upside versus the first-day close
Series H	Approx. $89.02	One of the latest private rounds	About 3.5x versus the first-day close
First-day secondary-market buyers	Near the highest opening price, closing price $311.07	Buy a scarce AI infrastructure name	Need both fundamentals and demand to absorb supply later

The meaning here is direct: secondary-market buyers are not only racing against future company value; they are also racing against the release schedule of early low-cost shares.

11.2 First Separate Four Types of Shares: Issuance, Float, Lock-Up Release, and Dilution

After an IPO, the easiest terms to confuse are "share count," "float," "lock-up release," and "dilution." These four words have different effects and cannot be placed in one basket.

Basis	Number or Rule	What It Affects	Common Misread
IPO issuance of Class A	30.000m shares	Main tradable supply on the first day	Mistakenly thinking this is the entire company share count
Post-IPO basic share count	215.110m shares, excluding overallotment	Basic market capitalization basis	Mistakenly thinking all of it can trade immediately
Overallotment option	Up to 4.500m shares	If exercised, increases cash, share count, and float supply	Mistakenly thinking it has necessarily already been exercised
Directed share program	Up to 1.500m shares, included in the 30.000m IPO shares	Changes the distribution of shares available to the general public	Mistakenly thinking it is an additional issuance
Employee early release	Up to 2.500m shares on the first trading day; up to another 2.500m shares on the second trading day	Earliest low-cost supply increase after the IPO	Mistakenly thinking there is no selling pressure during the lock-up period
Phased lock-up releases	Up to approx. 171.1m shares can be released in phases during the lock-up period	Changes the supply-demand balance over the next several months	Mistakenly thinking release occurs all at once only after 180 days
Potential dilution buckets	Options, RSUs, PSUs, OpenAI warrants, other warrants, plan reserves, etc.	Affects per-share value	Mistakenly thinking all of it equals current float

The key sentence is this: lock-up release is not dilution, and dilution is not necessarily immediately sellable. Lock-up release affects secondary-market supply; dilution affects per-share value. CBRS requires watching both.

11.3 First Stage: Low Float Amplifies Price Discovery

The main tradable supply on the first trading day was the 30.0m IPO shares, plus up to 2.5m early-release shares from non-executive employees. On the second trading day, if the first-day closing price is above 133% of the IPO price, up to another 2.5m shares can be released. 133% of $185 is $246.05; the first-day close of $311.07 appears to have satisfied the condition.

Time	Potential New Tradable Supply	Cumulative Early Supply	Trading Meaning
IPO first-day base supply	30.000m shares	30.000m shares	Main first-day trading pool, not the full company trading freely
First-trading-day employee early release	Up to 2.500m shares	32.500m shares	Low-cost employee shares may enter the market
Second-trading-day employee early release	Up to 2.500m shares	35.000m shares	After the first-day close exceeded $246.05, the condition appears satisfied
Overallotment option, if fully exercised	Up to 4.500m shares	39.500m shares	Increases cash, and also increases share count and supply

First-day trading volume was approximately equal to one full turnover of the 30m IPO shares, but far below the 215.110m basic share count. In other words, the $311.07 first-day closing price was formed under limited supply, not a fully traded equilibrium price for all potential shares.

11.4 Second Stage: Low-Cost Holders Begin Facing Liquidity Choices

A low cost basis does not mean holders will definitely sell, but it changes the psychological threshold for selling. Employees, early investors, and some option holders in particular may simultaneously face taxes, asset concentration, fund lives, risk controls, and rebalancing needs.

Round or Basis	Per-Share Cost or Reference Price	IPO Price / Cost	First-Day Close / Cost	Investment Meaning
Series A	Approx. $0.85	217.6x	366.0x	Extremely early cost basis; any release can carry enormous paper gains
Series B	Approx. $2.75	67.2x	112.9x	Early cost basis remains extremely low
Series C	Approx. $8.95	20.7x	34.8x	The gap versus secondary-market buyers is huge
Series D	Approx. $16.15	11.5x	19.3x	Mid-stage shareholders already have meaningful gains
Series E	Approx. $18.24	10.1x	17.1x	Mid-stage costs remain far below the IPO price
Series F	Approx. $27.74	6.7x	11.2x	Still meaningful paper gains versus the first-day close
Series F-1	$14.66	12.6x	21.2x	Low-cost round explicitly disclosed in the prospectus
Series G	$36.2324	5.1x	8.6x	One of the 2025 financing and tender reference prices
Series H	$89.0156	2.1x	3.5x	The latest private round, still clearly below the first-day close
Weighted-average exercise price of outstanding options	$4.97	37.2x	62.6x	Vested options are a source of potential supply and tax pressure

This cost gap determines a reality: for the secondary market to support a high price, it cannot rely only on "the company is good." It also has to absorb the gradual release of low-cost shares in the future.

11.5 Third Stage: The Lock-Up Period Does Not End All at Once, but Forms a Continuous Supply Curve

The 424B4 discloses that up to about 171.1m shares can be released in phases during the lock-up period, of which up to about 15.0m shares are held by directors and executives. This number is large, and the release does not wait until 180 days later to appear suddenly; it starts in phases from the early listing period.

Earliest Sale Time	Potential Released Shares	Cumulative Potential Release	Meaning for the Secondary Market
First trading day	2.5m	2.5m	Low-cost employee shares begin entering the market
Second trading day	2.5m	5.0m	After the first-day close satisfies the trigger condition, supply continues to increase
Second trading day after 2026 Q1 earnings release	27.7m	32.7m	The first large release window, combined with interpretation of the first post-listing earnings report
Second trading day after 2026 Q2 earnings release	36.4m	69.1m	One of the largest early release windows, testing demand depth
2026-08-19	14.6m	83.7m	Continuous release begins
2026-09-02	14.6m	98.3m	Continues testing absorption capacity
2026-09-16	14.6m	112.9m	Supply pressure continues
2026-09-30	19.4m	132.3m	Release size expands
2026-10-14	19.4m	151.7m	Release size expands
2026-10-28	19.4m	171.1m	Main releases during the lock-up period are close to complete
Lock-up period ends	Remaining shares	Undetermined	Earlier of: second trading day after 2026 Q3 earnings, or about 180 days after listing

This release schedule must be read together with valuation. The current price has already prepaid a high probability of success, while supply will continue to increase over the next several months. Only if fundamental evidence strengthens at the same time will the secondary market find it easier to absorb this supply curve.

11.6 Two More Types of Supply That Are Easy to Overlook

In addition to the main lock-up release schedule above, there is small but real supply from taxes and options.

Source	Quantity or Basis	Why It Matters
RSU tax-related selling	Up to approx. 0.2m shares around 2026-06-24; up to approx. 0.4m shares around 2026-08-18	Not active bearishness, but taxes needed after vesting
Vested options	Approx. 21.5m vested options as of 2026-04-30	Low exercise price may lead to share sales to cover exercise price and taxes
OpenAI warrants	Up to 33.445m Class N shares, vesting by conditions	The stronger the demand validation, the more denominator pressure must be included in per-share value
2026 equity incentive plan and employee stock purchase plan	Plan reserves do not equal issued shares, but represent future incentive capacity	Will affect future per-share value, not immediate float

These types of supply do not necessarily change the price on their own, but they make the market more sensitive: once earnings fall short of expectations or trading volume declines, low-cost shares and potential dilution will amplify stock-price pressure.

11.7 Governance Structure Also Affects the Position of Common Shareholders

Public shareholders buy Class A economic rights, but control remains highly concentrated among Class B holders. The 424B4 basis shows that after the IPO, Class B holders own about 85.3% of shares and about 99.2% of voting power. In other words, public shareholders mainly participate in economic outcomes, but have little ability to influence strategy, equity compensation, M&A, capital expenditures, and governance arrangements.

Item	Basis	Investment Meaning
Class A	1 vote per share	Listed trading shares mainly held by public shareholders
Class B	20 votes per share	Pre-IPO shareholders and insiders retain control
Class N	No voting rights, convertible into Class A	OpenAI warrants and other arrangements may enter the economic denominator
Class B share ownership	Approx. 85.3%	After the IPO, economic shares are still mainly held by original shareholders
Class B voting power	Approx. 99.2%	Public shareholders have very weak governance influence

This is not necessarily bad, but it requires investors to include "how management and large shareholders will allocate future value" in their judgment. Subsequent equity compensation, capital expenditures, customer warrants, and financing arrangements can all affect the per-share value common shareholders ultimately receive.

11.8 Secondary-Market Absorption Requires Watching Six Signals

Whether CBRS can absorb future lock-up releases does not depend only on whether the company has news. It depends on whether demand and supply match.

Signal	Good State	Bad State	Why It Matters
Trading volume	Volume expands around release dates but price remains stable	High-volume declines and weak rebounds	Judges whether new supply is absorbed by long-term capital
Price position	Can remain steadily above $185 and key pre-release prices	Falls below the IPO price or remains long below the early-release trigger price	Affects employee, institutional, and short-term capital behavior
Buyer composition	13F filings show long-term institutions, thematic funds, and potential index demand	Mainly relies on short-term capital and retail enthusiasm	Determines whether price absorption is durable
Fundamental synchronization	Earnings also provide evidence of revenue, gross margin, cash flow, and customer replication	Only story, without improvement in the three statements	The stronger the fundamentals, the easier lock-up releases are to absorb
Seller disclosures	Form 4 selling is explainable and the scale is absorbable	Insiders and low-cost holders continuously sell large amounts	Judges whether low-cost shares suppress valuation
Equity denominator	Diluted share count is explainable and per-share metrics improve	Warrants, RSUs, options, and plan reserves expand quickly	Company value growth may be offset by per-share dilution

The most important combination is: volume expands but the price does not break down, and earnings improve at the same time. Only this combination can show that the secondary market is not short-term enthusiasm, but real capital willing to absorb future releases at higher prices.

11.9 Three Trading Paths

Path	What Happens	Investment Meaning
Strong absorption	Earnings deliver, non-OpenAI customers increase, gross margin and cash flow improve; release-date volume expands but price remains stable	The market is willing to treat new supply as an entry opportunity, and the platform probability holds
Digestion period	Fundamentals advance, but evidence is not fast enough; lock-up and valuation pressure appear alternately	The stock may trade sideways or with wide volatility, using time to digest the high first-day valuation
Supply pressure	Lock-up supply increases, trading volume declines, and earnings do not simultaneously show gross-margin and cash-flow improvement	The market reprices it from a scarce asset to a high-expectation new issue, with multiple compression

The conclusion is direct: CBRS not only needs to prove that the company has a future; it also needs to prove that the stock itself can absorb new supply. Over the next several months, every earnings report, early release, lock-up release, Form 4, 13F, and volume change will tell investors whether the price around $311.07 reflects long-term capital recognition or short-term price discovery under low float.

---

Chapter 12: Second-Layer Validation Beyond OpenAI: API, Code Agents, Model Services, and Cloud Channels

OpenAI is the strongest anchor, but it cannot be the whole story. For Cerebras to move from a special-project supplier to a platform, it needs to prove four things: the API has sustained usage, code agents have paid retention, model services can be reused, and cloud channels do more than turn it into a hidden backend.

Second-Layer Validation	Positive Signal	Negative Signal
Inference API	Disclosure of usage, customer count, revenue, or gross margin; more production traffic migrates	It stays at benchmark and trial usage, without paid usage
Cerebras Code	Moves from sold out to a sustainable paid product	Sold out is only capacity control or a small-scale internal beta
AI Model Services	Customers repurchase, and project delivery is reusable	It becomes labor-intensive custom project work with unstable margins
AWS / HF / OpenRouter / Vercel	Brings net revenue, usage, and new customers	Only brings traffic, while the customer gateway and gross margin are captured by channels
Enterprise search / voice and video / research customers	Renewals, expansions, or harder commercial disclosure appear	Customer cases remain demos or small-scale experiments

If customer replication beyond OpenAI appears, Cerebras' revenue quality will migrate from "strong validation by a single large customer" toward "replicable platform." If no replication appears, the OpenAI contract still has value, but the valuation ceiling will be limited by customer concentration and bargaining power.

---

Chapter 13: Who Will Split the Money From This Low-Latency Inference Path

Cerebras' competition question should not be written as "there are many AI chip companies." The real question is: when customers decide to pay for low-latency inference, who ultimately keeps that money?

This money has three possible outcomes. First, customers directly buy Cerebras capacity or interfaces, and Cerebras receives revenue, gross profit, and cash flow. Second, customers use similar capabilities through cloud platforms, GPU clouds, or third-party APIs, while Cerebras is only a backend or is bypassed. Third, after absorbing the experience, large customers keep profit inside their own systems through in-house chips, custom chips, cloud-native accelerators, or default GPU systems.

For investors, the most dangerous scenario is not "low-latency inference demand does not exist." The more realistic danger is: demand exists, but profit does not stay with Cerebras.

13.1 One Map to See Where the Money Can Flow

Low-latency inference is not a one-way road where only Cerebras charges. After customer budgets emerge, they may flow to default GPUs, cloud vendors' in-house chips, low-latency APIs, custom chips, compute clouds, customer self-builds, and physical infrastructure. Cerebras' opportunity is to turn low-latency experience into dedicated capacity and interface revenue; the risk is that demand exists, but others take the profit.

Diversion Path	Representative Players	Pressure on Cerebras	First Signal to Watch
Default GPU path	Nvidia / AMD	Customers stay in the familiar ecosystem, compressing the low-latency premium	Non-OpenAI customers, pricing, gross margin
Cloud vendors' in-house chips	TPU / Trainium / Maia / MTIA	Cloud platforms keep chip-optimization gains internally	Cloud-channel revenue share, customer ownership
Low-latency API	Groq, etc.	Speed mindshare and developer gateways are split away	API usage, retention, unit revenue
Custom chips	Broadcom / Marvell / customers' in-house teams	Large customers may bypass external dedicated suppliers in the future	New contracts, renewals, price renegotiation
Compute cloud	CoreWeave, Oracle, Lambda, etc.	Customers buy available capacity and do not care about the underlying chip	Utilization, collections, capital expenditure efficiency
Physical infrastructure	Power, liquid cooling, data centers, networking	Costs and depreciation eat into revenue quality	Gross margin, CapEx, operating cash flow

Therefore, Cerebras is not only comparing chips with Nvidia. It is also competing against the easiest path for customers. The more important low latency becomes, the larger the budget pool; but customer gateways, delivery costs, and alternative paths determine how much of that budget can ultimately remain.

13.2 Nvidia: The Biggest Pressure Is Not Chip Speed, but Being the Default Choice

Nvidia's pressure on Cerebras is not just GPU performance, but default status. When customers buy Nvidia, they get GPUs, networking, full-rack systems, CUDA, libraries, developer experience, cloud-vendor procurement paths, and enterprise deployment habits. This combination means customers rarely need to explain "why choose Nvidia."

Cerebras' difficulty is the opposite. It needs customers to explain why they are not taking the default path, why low latency is worth a separate purchase, and why the migration cost and delivery risk of a new architecture are worth bearing.

Therefore, Cerebras does not need to prove it is "faster than Nvidia." It needs to prove that in code agents, real-time voice, long outputs, multi-step agents, and highly interactive enterprise search, the speed difference changes user behavior, and user behavior changes budget ownership. Only then does it have a chance to carve out an independent budget alongside Nvidia's default path.

13.3 AMD: A Second Source More Easily Accepted by Procurement

AMD's threat is more subtle. It does not need to prove itself as a completely new path; it only needs to prove it is a reliable second source within GPU workflows. For many customers, reducing reliance on Nvidia, getting better pricing, and increasing supply resilience are easier to pass through procurement and engineering review than adopting a completely new dedicated architecture.

This means that if the customer's real need is only "do not rely entirely on Nvidia," AMD is more natural than Cerebras. If the customer's need is "certain interactive tasks must be clearly faster," Cerebras has a clearer position.

Financially, this will show up in sales cycles and pricing. If Cerebras keeps facing customers that treat it as an ordinary alternative supplier and pressure pricing, rather than paying a premium for a low-latency experience, gross margin will be hard to support a platform valuation.

13.4 TPU, Trainium, and Maia: Cloud Vendors Keep Profit Internally

Google TPU, AWS Trainium, Microsoft Maia, and Meta MTIA represent the same direction: large customers are not only buying compute, they are also rewriting the compute supply chain. The purpose of cloud vendors' in-house chips is not to "sell chips," but to lower cloud service costs, bind customers, reduce external dependence, and turn hardware optimization gains into cloud gross margin.

This has two sides for Cerebras. The good side is that the industry has accepted non-GPU paths. The bad side is that the largest customers have the strongest motivation to turn non-GPU paths into their own internal capabilities.

The AWS partnership especially needs to be separated. Entering the AWS or Bedrock ecosystem can reduce distribution friction, but entering a cloud channel does not equal owning cloud customers. The customer gateway, billing relationship, service responsibility, and revenue-sharing rules will all determine whether Cerebras is taking platform profit or only serving as a dedicated backend behind the cloud platform.

13.5 Groq: Who Owns the "Speed" Mindshare?

The competition between Groq and Cerebras is not entirely a contest of hardware form factors. Both are competing for an easier-to-understand market position: low-latency inference.

Groq looks more like it enters from the API and developer gateway, first letting users feel the speed; Cerebras currently looks more driven by large-customer contracts and capacity buildout. The former is lighter and spreads quickly; the latter is heavier, but has stronger contract visibility.

If low-latency inference becomes an independent budget pool, both companies may benefit. But investors need to distinguish who owns the developer gateway, who owns large-customer contracts, and who can turn usage into gross profit and cash flow faster. Speed mindshare can be shared; profit does not necessarily have to be.

13.6 Broadcom, Marvell, and Customer Self-Builds: The Route for Large Customers to Bypass External Platforms

Broadcom and Marvell do not necessarily own end AI applications directly, but they are close to large customers' custom-chip and interconnect engineering. Their value lies in helping hyperscalers and model companies turn in-house ideas into manufacturable, deployable, connectable data-center infrastructure.

This path creates significant pressure for Cerebras. Customers such as OpenAI may need external low-latency capacity today, but if future workloads stabilize, scale becomes large enough, and cost pressure becomes strong enough, they will have incentives to self-build or customize part of the capability. At that point, Cerebras' role may shift from "long-term platform" to "transition-period supplier."

Therefore, the larger the OpenAI contract is, the more investors must ask two questions at the same time: how much demand does it prove, and does it also help the customer learn how to bypass you in the future?

13.7 Cloud Platforms and Compute Clouds: Customers Buy Available Capacity

Compute platforms such as CoreWeave, Oracle Cloud, Lambda, Crusoe, and Nebius remind us that AI infrastructure profit does not necessarily belong to the most special chip. It may also belong to the platform best at financing, procurement, go-live execution, scheduling, and signing customer contracts.

Many customers do not want to manage underlying hardware and do not want to judge which architecture is optimal. They only want stable, cheap, scalable compute that can be delivered on time. As long as GPU clouds or cloud platforms can optimize latency to "good enough," Cerebras' dedicated path must provide a stronger reason.

This is why the earlier three-statement analysis is important. Cerebras not only has to prove speed; it also has to prove capacity go-live, utilization, gross margin, operating cash flow, and per-share value can keep up together.

13.8 Which Signals Show Cerebras Is Defending Profit

Signal	Good Direction	Bad Direction	Why It Matters
Non-OpenAI customer revenue	Share rises and customer types diversify	Revenue is still defined by a few large customers	Proves low latency is not a single-customer special project
API and model-service usage	Calls, retention, and paying customers grow	Only trials and marketing heat	Proves Cerebras is getting the customer gateway
Cloud-channel economics	Gross margin is healthy after revenue share, and customer ownership is clear	Cloud platforms take the gateway and most of the profit	Determines whether it is platform revenue or backend supply
Gross margin	Stable or improving with scale	Compressed by depreciation, power, hosting, and reimbursed costs	Determines whether revenue can turn into profit
Operating cash flow	Improves with revenue growth	Maintained by loans, prepayments, or financing	Determines whether growth self-funds
New future revenue contract pool	Continues to increase beyond OpenAI	New additions still mainly come from one customer	Judges whether the company is moving from project to platform
Pricing and contract terms	Customers are willing to lock capacity and pay a premium	Customers renegotiate price or shorten commitments	Reflects pricing power and substitution pressure
Equity denominator	Per-share value still grows	Warrants, equity compensation, and financing dilution consume growth	Company value growth does not equal common-shareholder benefit

13.9 Which Situations Would Show the Money Is Being Taken by Others

The most dangerous situation is not that there is no demand for low-latency inference, but that demand exists and Cerebras cannot retain profit. Six types of signals need close observation:

Risk Signal	Which Link in the Value Chain It Breaks	First Metric to Watch
No replication beyond OpenAI	Demand validation -> platformization	Non-OpenAI revenue, customer count
Cloud channels have traffic but no profit	Distribution -> gross-margin ownership	Channel revenue share, service gross margin
Default GPU / cloud in-house paths catch up	Speed difference -> independent budget	Pricing, renewals, competitor performance
Large customers begin self-building or customizing	Long-term contract -> in-house replacement	New contracts, price renegotiation
Delivery costs consume the gains	Revenue -> gross profit -> cash	Depreciation, power, hosting, CapEx
Lock-up releases and dilution suppress per-share value	Company value -> per-share value	Diluted share count, Form 4, trading volume

The conclusion of Part Nine is this: Cerebras' opportunity comes from low-latency inference becoming an independent budget; its risk comes from that budget being jointly split by default GPUs, cloud vendors, in-house chips, API platforms, compute clouds, and physical infrastructure. The real validation is not "is it challenging Nvidia," but whether the next few quarters show more customers, healthier gross margin, steadier cash, and less per-share dilution.

Chapter 14: What the Current Price Has Already Prepaid

Valuation should not first ask "how large is this industry," but "what does the current price require the company to achieve." Cerebras' first-day trading already built many favorable future outcomes into the price in advance: the OpenAI contract turns into revenue on schedule, low-latency inference becomes an independent budget, gross margin is not damaged by delivery costs, operating cash flow can turn positive, customers beyond OpenAI can replicate, and equity dilution does not consume per-share value.

If these conditions are only partially met, the stock does not necessarily have to collapse, but the valuation method will change: the market will shift from "low-latency inference platform" to "OpenAI special-project supplier" or "high-growth but capital-heavy compute services provider."

14.1 First Separate Four Valuation Bases

Several numbers appeared on Cerebras' first listing day and must be used separately. The official IPO announcement gives the $185 offering price, 30m shares issued, and 4.5m-share overallotment option; media reports on first-day trading give the $311.07 closing price and about $66-67B basic market capitalization; TechCrunch also gives an approximately $56.4B fully diluted valuation at $185. They are not contradictory; they use different share-count denominators.

Basis	How It Is Calculated	Implied Value	Multiple of 2025 Revenue	What Question It Is Suited to Answer
IPO basic market capitalization	$185 × 215.110m basic shares	Approx. $39.8B	Approx. 78.1x	How much expectation the offering price already assigned to the company
First-day close basic market capitalization	$311.07 × 215.110m basic shares	Approx. $66.9B	Approx. 131.2x	How much the first-day secondary-market close prepaid
IPO fully diluted valuation, media basis	TechCrunch-disclosed fully diluted valuation corresponding to $185	Approx. $56.4B	Approx. 110.6x	After including potential dilution, the offering price is not cheap
First-day close implied fully diluted valuation	$311.07 × approx. 304.9m shares	Approx. $94.9B	Approx. 186.0x	If the market uses a broader equity denominator, the validation threshold is much higher

The point of this table is not precision to the decimal point, but a reminder: the same $311.07 creates completely different pressure under the basic share-count basis versus the fully diluted basis. Basic market capitalization looks about $67B, while the main dilution buckets bring it close to $95B. What investors truly buy is future diluted per-share value, not the company total value in a headline.

14.2 Once the Equity Denominator Changes, Price Pressure Is Completely Different

Using the first-day closing price of $311.07 as the base, the market-implied equity value and revenue multiples under different share-count bases are as follows.

Share-Count Basis	Shares	Corresponding Equity Value	Multiple of 2025 Revenue of 510m	Multiple of 2026-2027 Average Future Contract Revenue Conversion of 1.845B	Multiple of 2028-2029 Average Future Contract Revenue Conversion of 5.289B
Basic share count	215.110m	66.91B	131.2x	36.3x	12.7x
Near-term dilution	267.308m	83.15B	163.0x	45.1x	15.7x
OpenAI early trigger	272.882m	84.89B	166.4x	46.0x	16.0x
Main disclosed buckets	307.990m	95.81B	187.9x	51.9x	18.1x
All disclosed buckets, including plan reserves	354.194m	110.18B	216.0x	59.7x	20.8x

This reverse-engineering shows that Cerebras is not a case where "revenue growth is enough." If 2026-2027 annual average recognized revenue is about $1.845B, the stock would still trade at 36x revenue on the basic share-count basis and close to 52x revenue on the main disclosed-bucket basis. Even if annual average recognized revenue reaches about $5.289B in 2028-2029, the main disclosed-bucket basis is still about 18x revenue. A price like this requires the company not only to deliver revenue, but also to deliver high gross margin, better cash flow, customer replication, and platform probability.

14.3 Six Things the Market Has Already Prepaid

Assumption Already Prepaid	What the Price Roughly Believes	Evidence Needed	What Happens If Evidence Does Not Arrive
Rapid revenue expansion	The future revenue contract pool converts into revenue on schedule, and the 2026-2029 revenue step-up is clear	Future revenue contract pool roll-forward, revenue recognition, capacity go-live, delay disclosures	Revenue multiples compress, and the market stops pricing off the large contract total
Gross margin holds	Low-latency inference is not a low-margin delivery project, and gross margin after scale is not damaged by depreciation and hosting costs	Gross margin, service gross margin, depreciation, pass-through costs, data-center expenses	Reclassified from platform stock to capital-heavy services provider
Operating cash flow turns positive	Customer prepayments, collections, working capital, and capital expenditures can support self-funding	Operating cash flow, capital expenditures, receivables, contract assets, contract liabilities	Revenue growth is viewed as cash-consuming growth
Customer replication beyond OpenAI	Low-latency inference is not a single-customer special project	Non-OpenAI revenue, AWS/API usage, enterprise customers, model-service customers	Platform probability declines, and the valuation anchor returns to special-project supplier
Competition does not quickly pressure pricing	Nvidia, AMD, TPU, Trainium, Groq, and customer self-builds cannot quickly make low latency good enough	Customer migration cases, pricing, renewal terms, API retention, competitor performance	Speed premium compresses, and both gross margin and multiples are revised down
Dilution does not consume per-share value	OpenAI warrants, RSUs, options, plan reserves, and later financing do not offset company value growth	Diluted share count, warrant vesting, equity-compensation expense, revenue per share, cash flow per share	The company becomes larger, but common-shareholder per-share value does not

14.4 What Revenue Level Is Needed Before Valuation Starts to Look Less Tight

The table below is not a target price and not a forecast. It simply asks in reverse: if the market is willing in the future to give Cerebras 20x, 15x, 12x, 10x, or 8x revenue, how large must revenue become under different share-count bases to explain the price near $311.07?

Target Revenue Multiple	Revenue Needed Under Basic Share Count	Revenue Needed Under Near-Term Dilution	Revenue Needed Under OpenAI Early Trigger	Revenue Needed Under Main Disclosed Buckets	Revenue Needed Under All Disclosed Buckets
20x	3.35B	4.16B	4.24B	4.79B	5.51B
15x	4.46B	5.54B	5.66B	6.39B	7.35B
12x	5.58B	6.93B	7.07B	7.98B	9.18B
10x	6.69B	8.32B	8.49B	9.58B	11.02B
8x	8.36B	10.39B	10.61B	11.98B	13.77B

The reading is direct: under the main disclosed-bucket basis, for the price near $311 to look less tight, Cerebras probably needs revenue to cross the $5B-$10B level, and the market still has to be willing to give a high-growth AI infrastructure company a high revenue multiple. If revenue only reaches the 2026-2027 average contract conversion level, valuation remains crowded; if revenue can approach the 2028-2029 average contract conversion level, valuation pressure declines, but gross margin and cash flow still need to be proven.

14.5 The Gross-Margin and Cash-Flow Threshold Is More Important Than Revenue

The heavier the revenue, the less revenue multiples can be viewed alone. Cerebras has to build capacity, bear depreciation, pay power and hosting costs, and face warrants and equity compensation. The market will ultimately shift from revenue multiples to free cash flow and cash flow per share.

Annual Revenue Level	35% Gross Margin	40% Gross Margin	45% Gross Margin	50% Gross Margin	55% Gross Margin
1.845B	0.65B	0.74B	0.83B	0.92B	1.01B
5.289B	1.85B	2.12B	2.38B	2.64B	2.91B
8.000B	2.80B	3.20B	3.60B	4.00B	4.40B
10.000B	3.50B	4.00B	4.50B	5.00B	5.50B

This means that if 2028-2029 average annual revenue is close to $5.3B, even a 45% gross margin produces only about $2.4B of gross profit. After deducting R&D, sales, G&A, operating costs outside depreciation, working capital, and capital expenditures, free cash flow still has to be discounted again. If the equity denominator is viewed under the main disclosed buckets or all disclosed buckets, what the market needs is not "there is gross profit," but "gross profit is high enough, capital expenditures are controlled enough, and free cash flow comes fast enough."

14.6 How Much Free Cash Flow Is Needed for the Stock to Look Like a Mature Asset

If investors in the future are no longer willing to price only by revenue multiples and instead require a free cash flow yield, the cash-flow threshold corresponding to $311.07 will be very high.

Required Free Cash Flow Yield	FCF Needed Under Basic Share Count	FCF Needed Under Near-Term Dilution	FCF Needed Under OpenAI Early Trigger	FCF Needed Under Main Disclosed Buckets	FCF Needed Under All Disclosed Buckets
2%	1.34B	1.66B	1.70B	1.92B	2.20B
3%	2.01B	2.49B	2.55B	2.87B	3.31B
4%	2.68B	3.33B	3.40B	3.83B	4.41B
5%	3.35B	4.16B	4.24B	4.79B	5.51B

In 2025, the company's operating cash flow was about -10m, capital expenditures were about 382.7m, and rough free cash flow was about -392.7m. The gap versus the table above is clear: the market today is not buying an already mature cash-flow asset, but the probability that it may become a high-cash-flow platform in the future.

14.7 Platform Probability Threshold: The Price Buys Not Only Revenue, but an Identity Conversion

The most expensive part of Cerebras' current price is not 2026 revenue or 2028 revenue, but "identity conversion." The market is assigning it some platform probability: from an OpenAI-driven dedicated capacity supplier into a replicable, callable, long-term low-latency inference platform.

Future Identity	How the Market Would Price It	Conditions That Need to Hold	Valuation Consequence If They Do Not Hold
Special-project supplier	Priced by large-customer projects, delivery risk, and cash consumption	OpenAI contract advances, but customer concentration remains high and non-OpenAI replication is weak	Revenue multiple falls sharply, and the stock mainly follows project milestones
Capital-heavy compute services provider	Priced by revenue growth, utilization, depreciation, and cost of capital	Capacity can go online and customers can expand, but capital expenditures remain heavy long term	Multiple is below asset-light platforms, and cash-flow validation is critical
Dedicated low-latency inference platform	Priced by customer gateways, API retention, gross margin, and network effects	Customer replication beyond OpenAI, interface usage growth, gross-margin and cash-flow improvement	Can support a higher revenue multiple, but still has to bear dilution and competition
One default path for AI infrastructure	Priced as a scarce platform asset	Low latency becomes a major budget category, and Cerebras becomes one of the default multi-customer choices	This is the most optimistic layer in the current price, and the evidence is still insufficient

Therefore, the conclusion from the valuation reverse-engineering is not "it is expensive, so it cannot be studied," but "the price has already counted the identity conversion in advance." If subsequent evidence proves only that the OpenAI contract is real, that is still not enough. It must also prove that this contract can bring multi-customer replication, platform gateways, stable gross margin, operating cash flow, and per-share value.

14.8 Three Repricing Paths

Path	What Needs to Happen	Stock-Price Meaning	What Investors Should Watch Most
Continued upward revision	The future revenue contract pool converts to revenue on schedule, non-OpenAI customers increase, gross margin is stable, operating cash flow improves, and the equity denominator is controlled	The market continues to accept platform probability, and may even tolerate high short-term multiples	RPO roll-forward, customer concentration, service gross margin, OCF, diluted shares
High-level digestion	Revenue advances, but evidence on cash flow, customer replication, and dilution is not fast enough	The stock fluctuates around earnings and lock-up releases, while valuation digests over time	Trading volume, lock-up supply, gross profit, CapEx, API usage
Downward repricing	Delivery delays, gross-margin pressure, insufficient replication beyond OpenAI, or competing paths compress the speed premium	Repriced from platform stock to project supplier or capital-heavy services provider	Delay disclosures, gross margin, customer structure, contract renegotiation, operating cash flow

The most prudent positioning is this: Cerebras is a candidate winner in low-latency inference platforms, with strong customer validation and strong revenue visibility, but the current price has already prepaid a lot of success. It is no longer an undiscovered name, but a public-market AI compute infrastructure asset with high expectations, high volatility, and high validation pressure.

14.9 Historical Post-IPO Performance: A First-Day Jump Does Not Equal Long-Term Outperformance

After CBRS's first-day surge, the stock price cannot be explained only by "company quality," and the company cannot be dismissed only by "first-day heat." Returns after a new issue split into two layers: IPO allocation investors calculate from the $185 offering price, while secondary-market investors calculate from the first-day open, first-day close, or post-listing high. The two are completely different.

Jay Ritter / University of Florida's long-term IPO return data provide a baseline: from 2001 to 2024, U.S. operating company IPOs had average first-day returns of 18.8%, but 3-year market-adjusted returns of -14.7%; VC-backed tech IPOs had higher average first-day returns of 54.3%, but 3-year market-adjusted returns of -19.6% and style-adjusted returns of -26.6%.

This historical data is not a forecast, but a reminder: first-day pop is normal in the IPO market, but chasing after the first day has not been favorable on long-term averages. CBRS's first-day +68.15% is already above the ordinary IPO average and above the long-term sample mean for VC-backed tech IPOs. For IPO allocation investors, this is a huge gain; for those buying near the first-day closing price, it is a higher fundamental validation threshold.

CBRS will have to pass three exams at the same time going forward:

Exam	What the Market Will Watch	Failure Signal
New-issue supply-demand exam	Trading volume, early releases, lock-up period, absorption of low-cost shares	The stock price can only be maintained by low float
AI infrastructure exam	Customers beyond OpenAI, capacity go-live, data-center delivery	The contract pool is large but delivery is slow
Public-company earnings exam	Revenue, gross margin, operating cash flow, CapEx, diluted share count	Revenue growth does not turn into cash, and per-share value is diluted

Chapter 15: Next-Quarter Review Checklist

Starting next quarter, the key review method is to put every quarter back onto the same path: whether capacity goes online, whether the future revenue contract pool converts into revenue, whether gross margin is stable, whether cash improves, whether share count is controlled, and whether the price can still absorb new supply.

Forecast Item	Current Baseline	Positive Outcome	Negative Outcome	Impact on Judgment
OpenAI capacity go-live	Initial 750MW commitment, entering delivery schedule from 2026	Site, power, cooling, hardware, and acceptance schedule are clear	Delivery delays, vague disclosure, or responsibility shifts	Determines whether the future revenue contract pool can enter the income statement
Future revenue contract pool converts into revenue	Approx. $24.6 billion future revenue contract pool is the largest source of revenue visibility	The portion recognizable within one year rises, and revenue recognition timing matches contract expectations	Contract pool shifts later, shrinks, is renegotiated, or is unclearly explained	Determines whether the market continues to assign high revenue multiples
Gross margin	Still needs continuous validation	Revenue grows while gross margin remains stable or improves	Pass-through and reimbursed costs, power, hosting, depreciation, and operations consume gross profit	Determines whether this is platform-style revenue or project-style revenue
Cash flow	Operating cash flow, capital expenditures, and working capital remain weak points	Operating cash flow improves, capital expenditures / revenue decline, and working-capital pressure eases	The more revenue grows, the larger cash consumption becomes	Determines whether growth continues to rely on external capital
Customer replication	OpenAI is strong validation, but customer concentration is clear	AWS / API / enterprise customer revenue grows, and OpenAI share declines	No substantive progress among non-OpenAI customers	Determines whether it is a replicable platform or a special-customer project
Per-share value	Basic shares, OpenAI warrants, RSUs, options, and lock-up releases all affect the denominator	Revenue and cash grow faster than dilution	Dilution pace offsets operating improvement	Determines how much common shareholders ultimately retain
Share supply	First-day close of $311.07, with future early releases, overallotment, and phased lock-up releases	Under high trading volume, the price can absorb new supply	High-volume declines around lock-up releases	Determines whether short- and medium-term entry points are damaged by supply-demand dynamics

Product and Customer Evidence Validation

Future quarters also need more attention to several product signals. Product existence is not the endpoint; real payment, retention, gross margin, and customer replication are the endpoint.

Product or Customer Clue	What Should Be Seen	What Should Not Mislead
Inference API	Disclosure of usage, customer count, revenue, or gross margin	Only looking at tokens/sec or benchmarks
Cerebras Code	Moves from sold out to a sustainable paid product	Misreading capacity control or a small-scale internal beta as a revenue breakout
AI Model Services	Forms replicable services, rather than one-time engineering projects	Treating customer logos as revenue proof
AWS, Hugging Face, OpenRouter, Poe, and other channels	Bring net revenue, usage, and new customers	Only looking at number of integrations, not customer gateways and gross-margin ownership
Cases such as Notion, AlphaSense, Cognition, Tavus, OpenCall, and Sei	Renewals, expansions, or harder commercial disclosure appear	Treating demo cases as long-term contracts
White-paper benchmarks	Reviewed by third parties and production workloads	Treating performance under specific methodologies as general unit economics