Signals Inbox·July 28, 2026·AI Chips

Who is Nvidia’s biggest startup threat?

Cerebras is Nvidia’s biggest startup threat today: still tiny beside the GPU giant, but already turning specialist inference speed into paid workloads, hyperscale commitments and a real deployment business.

Q1Why has Nvidia’s startup threat become a real question now?

Nvidia now has a credible startup problem because AI inference is breaking into separate jobs, and specialist chips are starting to win real production roles.

A large AI request usually has two main stages. During “prefill,” the system reads and processes the prompt. During “decode,” it generates the answer one token at a time. GPUs can handle both, but the two stages place different demands on the hardware. Prefill rewards heavy parallel processing, while decode often depends more on memory speed and rapid sequential calculation.

Companies are now designing systems around that difference. AWS plans to use Trainium processors for prefill and Cerebras systems for decode. AMD has announced a similar setup combining its Helios platform with Cerebras’s Wafer-Scale Engine. AMD said the joint system could deliver up to five times more tokens per second per watt than the comparison system used in its testing.

OpenAI has made the largest commitment so far, agreeing to deploy 750 megawatts of Cerebras inference capacity in stages. That is far beyond a laboratory trial or a small cloud listing. It gives Cerebras a path into one of the world’s largest AI workloads.

Nvidia’s own response tells us the specialist-chip opportunity is real. It licensed Groq’s inference technology, hired several of Groq’s key leaders and added a Groq-derived LPX processor to the Vera Rubin platform. Nvidia now plans to pair GPUs with a processor designed specifically for fast, sequential token generation.

Three different groups have therefore reached a similar conclusion: AWS and Cerebras, AMD and Cerebras, and Nvidia with Groq technology. These systems keep general-purpose processors for some tasks while handing latency-sensitive decoding to more specialized hardware.

That shift gives startups a realistic opening. They no longer need to replace Nvidia across training, networking, simulation and every form of inference. One expensive job where a different architecture works much better is enough to get started.

Q2What would a startup have to do to truly threaten Nvidia?

A startup threatens Nvidia when it repeatedly moves valuable workloads away from Nvidia hardware and can keep doing so at scale.

A faster benchmark alone tells us very little. Chip companies can choose a favorable model, batch size, context length or latency target and produce a spectacular result. The harder test begins when customers run the hardware continuously, pay for more capacity and expand after seeing the real costs.

A challenger can create serious pressure without replacing Nvidia everywhere. Capturing a large share of real-time inference would be enough. That would reduce the number of Nvidia GPUs required for serving models, challenge Nvidia’s pricing and weaken the assumption that every important AI workload belongs on CUDA.

Independence counts too. Groq’s architecture remains technically relevant, but Nvidia now licenses the technology and includes it in its own systems. GroqCloud can still compete for customers, yet Groq’s core idea also strengthens Nvidia.

We judge each startup on five practical tests.

Five tests for a credible Nvidia challenger

Q3Why is Nvidia still so hard to challenge?

Nvidia remains overwhelmingly difficult to challenge because customers buy a complete computing system, while most startups still offer one unusually strong component.

In Nvidia’s latest reported quarter, the company generated $81.6 billion in revenue. Its Data Center business alone produced $75.2 billion, including roughly $60.4 billion from compute and $14.8 billion from networking.

The networking figure shows why a simple chip comparison misses so much. Nvidia earned more from data-center networking in three months than any independent AI-chip startup has generated in its entire history. It can sell processors, switches, interconnects, server designs and software as one package.

A customer building a large AI cluster also needs compilers, optimized model libraries, debugging tools, monitoring, reliable communication between thousands of processors and engineers who know how to use everything. Nvidia has spent nearly two decades building that environment around CUDA.

A startup can produce a faster chip and still lose the sale because the buyer may then spend months adapting models, replacing software, retraining engineers and fixing reliability problems.

Nvidia can also improve its GPUs, add a specialist processor, adjust pricing or bundle more software and networking into the deal. The Groq licensing agreement shows how quickly it can respond when an alternative architecture becomes useful.

Q4Which AI-chip startup has built the largest real business?

Cerebras currently has the strongest commercial proof among independent AI-chip startups.

Its latest published quarterly filing showed $193.4 million in revenue, up 94% from the same period a year earlier. Hardware contributed $110.6 million, while cloud and services revenue reached $82.8 million after growing 178%.

The quality of those results improved too. Gross margin rose from 41.8% to 44.6%, and the company’s net loss narrowed to $14 million. Cerebras remains unprofitable, but the loss is modest compared with its revenue and far smaller than many heavily funded chip startups at a similar stage.

The balance sheet gives Cerebras room to expand. At the end of the reported quarter, it held around $3.26 billion across cash, restricted cash and short-term investments. Its next quarterly results have been scheduled but are not public yet, so these remain the freshest confirmed financial figures.

The comparison with other startups is imperfect because most disclose usage, funding or contracts instead of audited revenue. Groq says it serves more than five million developers across 13 data centers and processes trillions of tokens each week, although it has not published comparable sales figures. Etched says production has begun for more than $1 billion in customer contracts. Tenstorrent has made its Galaxy Blackhole systems generally available. SambaNova says it finished the previous year with record bookings and revenue, without giving the amounts.

Cerebras is the only candidate that currently combines public financial statements, hundreds of millions of dollars in annualized revenue, expanding cloud sales and hyperscale capacity commitments.

Leading AI-chip startups compared

Q5Is Cerebras actually taking business from Nvidia today?

Cerebras is already taking some inference work that would otherwise be a natural fit for Nvidia, although the amount remains tiny beside Nvidia’s business.

OpenAI’s commitment is the clearest example. The initial agreement covers 750 megawatts of Cerebras capacity, and Cerebras has already begun recognizing revenue from the arrangement. Its latest filing recorded $16.9 million of quarterly revenue connected with OpenAI.

AWS offers another route into Nvidia’s territory. Its planned system will use Trainium for prefill and Cerebras for decode, with access through Amazon Bedrock. Customers will be able to use Cerebras through an existing AWS relationship instead of procuring unfamiliar hardware themselves.

AMD’s new Cerebras partnership follows the same logic. AMD Helios will process prompts and large context windows, while Cerebras will generate tokens. The planned service will initially run through Cerebras Cloud.

These deals give Cerebras a defined role inside larger AI systems. That is an easier sale than asking a customer to replace its entire Nvidia setup.

The scale comparison remains brutal. Nvidia’s latest quarterly Data Center revenue was almost 389 times Cerebras’s total quarterly revenue. Cerebras can take meaningful workloads from Nvidia while barely affecting Nvidia’s current financial results.

The near-term pressure will appear in customer behavior rather than market share. Watch whether OpenAI expands beyond committed capacity, whether AWS customers choose the combined system repeatedly and whether AMD turns the technical partnership into a widely used product.

Nvidia and Cerebras at current scale

Q6Does Cerebras’s wafer-scale chip give it a lasting edge?

Cerebras has a genuine hardware advantage in fast token generation, but its lasting value depends on how often customers will pay for extreme speed.

Conventional chipmakers cut many processors from one silicon wafer. Cerebras builds one processor across almost the whole wafer. Its WSE-3 contains around four trillion transistors and 900,000 AI-focused cores.

The large design keeps much more computation and memory movement inside one device. That reduces the delays created when a model is divided across many separate processors that constantly exchange data.

This approach works especially well during decode. Each new token depends on the tokens generated before it, so simply adding more parallel processors eventually produces diminishing returns. Cerebras can keep the model close to enormous amounts of on-chip memory and generate each next token quickly.

The company has reported speeds approaching or exceeding 1,000 tokens per second on several large models. Some of those results have been measured by Artificial Analysis, which gives them more weight than an internal benchmark alone.

That speed is valuable for voice assistants, coding agents and interactive research tools. It matters much less for overnight batch processing.

Flexibility still favors Nvidia. GPUs can train models, serve many types of inference, run simulations, process video and support thousands of existing applications. Cerebras covers a narrower range of workloads.

Nvidia is already answering the specialist threat through LPX. Cerebras has a real advantage today, but calling it a permanent moat would be premature.

Q7Is the OpenAI deal a breakthrough or a dangerous dependency?

The OpenAI deal transforms Cerebras’s prospects, while also making its future heavily dependent on one customer.

The agreement covers 750 megawatts of committed capacity. Cerebras’s SEC filing also revealed that OpenAI can request another 1.25 gigawatts, which would bring the potential total to two gigawatts.

The financial commitment is enormous compared with Cerebras’s current size. The company reported $25 billion in remaining performance obligations, with a large share tied to the OpenAI agreement. That backlog equals almost 32 times Cerebras’s latest annualized quarterly revenue.

The revenue will arrive slowly. Cerebras expects to recognize around 16% of the backlog during the first 24 months, another 45% during months 25 to 48 and the rest later. A $25 billion backlog cannot be treated as $25 billion of near-term sales.

OpenAI is also helping finance the expansion. It provided Cerebras with a secured working-capital loan of roughly $1 billion carrying a 6% interest rate. Cerebras can repay parts of the loan through compute services and other agreed credits.

The arrangement includes warrants that vest when Cerebras reaches deployment and commercial milestones. OpenAI gains more than computing capacity from the relationship. It receives financial protection and potential equity upside.

Cerebras already had a concentration problem. In its latest quarter, MBZUAI represented 63% of revenue and G42 represented 11%. Three customers accounted for 86% of accounts receivable.

OpenAI improves the quality and geographic diversity of the customer base, but Cerebras is still built around a handful of very large relationships. A delay or renegotiation involving one of them could reshape an entire quarter.

Q8Can Cerebras deliver the capacity it has promised?

Cerebras has enough money and manufacturing support to begin the build-out, but delivering hundreds of megawatts is now its hardest test.

The company and Flex recently expanded their manufacturing partnership in California. New production lines, testing equipment and additional floor space are expected to support a sevenfold increase in CS-3 production capacity.

Cerebras is also building a European footprint. It plans to bring its first new European data-center capacity online before the end of the year and reach 200 megawatts across France and the Nordic region by the end of 2027. Part of that capacity will serve OpenAI.

The numbers show how quickly the company must grow. Its planned European footprint alone will require large numbers of specialized systems, high-power electrical equipment, liquid cooling, optical networking and trained technicians. The 750-megawatt OpenAI commitment is nearly four times larger.

Its financial filing revealed another useful detail: Cerebras had already signed data-center leases with around $2.3 billion in future minimum payments before those leases had even begun. That is a real industrial commitment, accompanied by substantial fixed costs.

The manufacturing process also looks harder than assembling conventional servers. Each CS-3 needs custom handling, cooling, calibration and full-system testing because the wafer-scale processor behaves differently from an ordinary accelerator board.

Cerebras can probably build enough systems to begin fulfilling the contracts. Whether it can deliver them quickly, reliably and without crushing its margins is the question that now decides the story.

Q9Did Nvidia neutralize Groq, and can it do the same to other challengers?

Nvidia has sharply reduced Groq’s value as an independent threat, but repeating the same move with every dangerous startup would be much harder.

The companies signed a non-exclusive licensing agreement covering Groq’s inference technology. Founder Jonathan Ross, president Sunny Madra and other employees joined Nvidia, while Groq continued operating separately.

Nvidia moved quickly from licensing to product development. The Vera Rubin platform now includes Nvidia Groq 3 LPX, a processor aimed at low-latency, large-context inference. Nvidia says the combined Rubin and LPX system can deliver up to 35 times more inference throughput per megawatt for selected workloads.

Customers attracted to Groq-style sequential processing may soon get it as part of an Nvidia rack, connected to Nvidia GPUs, networking and software. That closes one of the clearest gaps in a GPU-only system.

Groq itself remains active. The company recently raised another $650 million, operates 13 data centers and plans to move toward 200 megawatts of capacity by the end of 2027. It says more than five million developers use the platform.

Its strategy has changed, though. Groq’s newest expansion plan explicitly includes Nvidia LPX systems alongside its existing infrastructure. The company is becoming a large inference-cloud operator that uses technology shared with Nvidia.

GroqCloud could still win substantial business. Yet Groq’s success no longer automatically weakens Nvidia because part of that success can increase demand for Nvidia’s LPX product.

The deal was unusually convenient for Nvidia. It gained the technology and key employees without completing a full acquisition.

A similar arrangement would be harder with Cerebras. Cerebras is publicly traded, has a large balance sheet, major customer commitments and a commercial identity built around providing an alternative to GPUs. An Nvidia acquisition could also undermine the diversification that attracts customers such as OpenAI, AWS and AMD.

Etched could become a more plausible target while its systems are still entering production. Tenstorrent’s licensable intellectual property could also attract a larger semiconductor company.

Regulators would examine any deal involving a serious AI-compute challenger. Nvidia already controls most of the accelerator market, and buying a company capable of reducing that dominance would invite scrutiny.

Nvidia can still license useful technology, hire engineers, copy a system design or reduce a startup’s advantage through a new product. It handled Groq skilfully. Cerebras would be much more difficult to pull inside its platform.

Q10Could Etched overtake Cerebras as Nvidia’s biggest startup threat?

Etched could overtake Cerebras, but today it remains a high-upside manufacturing bet with almost no public operating history.

Etched’s Sohu processor is designed specifically for transformer models. By giving up some of the flexibility of a general GPU, the company expects to fit much more transformer computation into each chip and deliver better inference economics.

The company’s latest production update was unusually concrete. Etched said it had received its first A0 silicon, begun production for more than $1 billion in customer contracts and planned to ship its first racks during the summer.

It has built more of the surrounding operation than most early chip startups. Etched opened a factory in Taiwan and added a data center, test house and prototype-manufacturing laboratory at its San Jose site. These investments suggest that the company understands how much work sits between receiving a chip and delivering a dependable rack.

The missing evidence is still substantial. We have no recognized revenue, customer deployment results, uptime record or independent confirmation that the final systems deliver the promised economics.

Its specialization creates another risk. Transformer models dominate AI today, but hardware development takes years. Changes in model architecture, memory use, sparsity or inference techniques could reduce the advantage of a highly fixed design.

Etched has a plausible route to becoming dangerous very quickly. A few large customers expanding after the first deployments would change its position. Until then, Cerebras has more commercial proof, more production experience and a larger financial base.

Q11Can Tenstorrent weaken Nvidia by making AI hardware more open?

Tenstorrent currently poses a slower, broader threat because it attacks Nvidia’s lock-in through open software, standard networking and licensable processor designs.

The company has made its Galaxy Blackhole systems generally available, with entry pricing around $110,000. It sells complete machines, cloud capacity and intellectual property that other companies can use in their own chips.

That licensing model makes Tenstorrent unusual. A carmaker, electronics company, cloud provider or government project can embed Tenstorrent technology inside a product instead of buying every accelerator directly from Tenstorrent.

Its latest system also makes aggressive performance claims. Tenstorrent says Galaxy Blackhole can produce more than 350 tokens per second per user on a large DeepSeek model and generate certain AI videos ten times faster than leading GPU systems. It also says about 90% of models from Hugging Face can run on the platform.

Those numbers come from Tenstorrent and its partners, so we treat them as evidence of technical progress rather than settled market proof. The company has disclosed very little about system sales, recurring cloud usage or customer expansion.

Open software can also be harder for ordinary customers. Greater control appeals to expert engineering teams, while many buyers prefer a polished proprietary system that works immediately and comes with one company responsible for every layer.

Tenstorrent may eventually weaken the industry’s dependence on Nvidia by spreading an alternative architecture through many products. That route takes longer than selling fast inference to OpenAI, AWS or AMD. It could still become powerful if a few large manufacturers adopt the designs across millions of devices or servers.

Q12Is SambaNova closer to threatening Nvidia than it looks?

SambaNova looks more credible than its public profile suggests, although its newest hardware still has to prove itself in customer data centers.

The company has survived long enough to release five generations of its dataflow architecture. Its new SN50 processor targets agentic inference, where one user request may trigger many model calls and small delays accumulate quickly.

SambaNova says the SN50 can reach five times the maximum speed and three times better total cost of ownership than competing GPU systems under selected conditions. It also raised more than $350 million to expand manufacturing and cloud capacity.

SoftBank will be the first major customer. It plans to deploy SN50 systems in Japanese data centers for low-latency enterprise inference. The relationship carries more weight because SoftBank already hosts SambaCloud and has experience with the previous platform.

Intel is also involved through a planned multi-year collaboration covering manufacturing and cloud-scale infrastructure. That can give SambaNova access to a much larger supply chain and distribution network.

The weak point is timing. SN50 customer shipments are expected later this year, so most of the performance and cost claims still describe a product entering production. The SoftBank order has no disclosed size, and SambaNova does not publish audited revenue.

SambaNova belongs among the serious challengers because it has working technology, financing and a large first customer. Cerebras stays ahead because we can already measure its revenue, margins, contracted capacity and manufacturing expansion.

Q13Do startup benchmark wins against Nvidia hold up in production?

Startup benchmark wins are useful, but they rarely settle the buying decision.

Inference performance changes with the model, context length, batch size, number of simultaneous users, numerical precision and utilization rate. A system that produces tokens fastest for one user may become less attractive when thousands of requests arrive together.

The customer pays for more than accelerator time. Servers may need extra CPUs, memory, networking, cooling and engineering. A chip with a lower headline cost can become expensive when it sits idle or requires a separate cluster for unsupported models.

Software support causes similar problems. Cerebras, Groq, Tenstorrent and SambaNova all support popular open models, but Nvidia offers a much wider range of optimized applications. A new model architecture can reach Nvidia quickly because so many developers already build around CUDA.

Independent testing improves confidence, although it still captures only one configuration. Cerebras’s results on Artificial Analysis carry more weight than a company-selected demonstration. Production expansion carries more weight again because the buyer has seen the latency, reliability and full cost.

We give the most credit to three kinds of evidence: repeat capacity purchases, independently measured results and revenue from recurring usage.

How we weigh competitive evidence

Q14Can any startup get around Nvidia’s CUDA advantage?

No startup can currently match CUDA across the whole market, so the practical route is to hide the hardware behind a familiar cloud service or API.

Most developers do not want to redesign an application around a new processor. They want to send the same request to another provider and receive a faster or cheaper answer.

Cerebras and Groq both offer APIs designed to resemble OpenAI’s interface. A developer can often change an endpoint and a few settings instead of rewriting the entire application. The chip becomes almost invisible.

Cloud distribution pushes the idea further. AWS plans to make the Trainium and Cerebras system available through Amazon Bedrock. Customers can stay inside the AWS environment they already use for billing, security and deployment.

This avoids a direct fight with CUDA. The customer does not need to choose a new programming ecosystem because the cloud provider handles the hardware behind the service.

The limitation appears when customers need unusual models, custom operators, training or deep control over the system. Nvidia’s mature tools become much harder to avoid in those cases.

Startups have their best chance in standardized inference. The model is already trained, the request arrives through an API, and the customer mainly cares about speed, reliability and price.

Cerebras currently leads this route because it combines a simple cloud product with planned distribution through AWS and direct capacity for OpenAI.

Q15Are Amazon and Google bigger threats to Nvidia than the startups?

Amazon and Google are larger strategic threats to Nvidia than any startup because they can move their own enormous workloads onto chips they control.

AWS already offers Trainium3 systems for training and inference. A fully configured Trainium3 UltraServer can contain 144 chips, and AWS can connect large numbers of those systems inside its own data centers.

Google’s latest TPU generation, Ironwood, supports both training and inference. Its architecture can scale to pods containing as many as 9,216 processors, surrounded by Google’s own networking, storage and software.

These companies begin with advantages no startup has. They own global data centers, established cloud customers and major internal AI workloads. They can deploy a custom processor inside their own services before persuading outside companies to use it.

They can also price the hardware as part of a broader cloud relationship. A lower chip margin may be acceptable if the customer spends more on storage, databases and other services.

Their main weakness is customer choice. Some companies worry about becoming too dependent on one cloud provider. Nvidia hardware is available through almost every major cloud and server manufacturer, which makes it easier to move workloads between suppliers.

Startups can still supply specialist technology that the hyperscalers do not build themselves. AWS’s Cerebras partnership is the clearest example.

Q16What would prove Cerebras has become a major Nvidia threat?

Cerebras becomes a major Nvidia threat when customers expand after using it, margins hold during the build-out, and revenue spreads beyond three dominant relationships.

The first test is delivery. Cerebras needs to convert factory capacity, data-center leases and customer commitments into working systems. Deployed megawatts, uptime and utilization count more than announced plans.

The second test is repeat buying. OpenAI has already made a huge commitment, but the strongest evidence would be an expansion after the first capacity has run in production. The same applies to AWS, AMD and CrowdStrike. Follow-on orders show that customers have measured the economics and want more.

The third test is diversification. Cerebras needs meaningful revenue from customers outside OpenAI, MBZUAI and G42. Ten customers contributing 5% to 10% each would create a healthier business than three customers controlling most sales and receivables.

The fourth test is profitability. The latest 44.6% gross margin is encouraging, but new data centers bring rent, power, depreciation and staffing costs. Cerebras must show that cloud inference remains attractive after those expenses appear at full scale.

The fifth test is resilience against Nvidia’s response. Rubin and LPX will compete directly for low-latency inference. Cerebras will need to preserve a clear advantage in speed, cost or ease of deployment after those products become widely available.

We would consider the threat major once several independent customers expand deployments, annual revenue reaches several billion dollars and Cerebras can fund growth without relying on one customer’s financing.

Until then, it is the most credible startup challenger and still a small company taking on an enormous platform.

Q17Who is Nvidia’s biggest startup threat?

Cerebras is Nvidia’s biggest startup threat today because it has crossed the line from a clever chip to a real, financed deployment business.

It has the strongest combination of recognized revenue, audited financial information, large customer commitments, manufacturing expansion and cloud distribution. Its newest partnerships with AMD and CrowdStrike add fresh evidence that the company can reach workloads beyond its original government and research customers.

Etched has greater surprise potential. Its transformer-specific processor could deliver exceptional economics, and more than $1 billion in contracts gives it a serious starting point. We still need to see shipping systems, recognized revenue and customers expanding their deployments.

Tenstorrent attacks a broader weakness in Nvidia’s position by promoting open software, standard networking and licensable designs. That strategy could become influential over time, particularly in sovereign AI and custom chips, but it has not yet produced Cerebras-level commercial evidence.

SambaNova has strong technology and a credible first SN50 customer in SoftBank. The newest hardware is still entering production, which keeps it below Cerebras for now.

Groq developed one of the most strategically important alternative architectures. Nvidia’s licensing agreement and LPX product changed the competitive result. GroqCloud can keep growing, while Groq-derived technology also helps Nvidia.

Cerebras threatens only one part of Nvidia’s market. Its best position is ultra-fast decode for coding agents, voice systems, research tools and other applications where every second affects the product. Nvidia remains far stronger in training, broad inference, networking, software and general accelerated computing.

That limited scope is enough to matter. If Cerebras makes specialist inference a large independent category, Nvidia will lose some workloads, face more pricing pressure and share AI data centers with other architectures.

Our judgment is clear: Cerebras is the strongest independent startup threat, Etched is the most dangerous unproven challenger, and Nvidia has already turned Groq’s breakthrough into part of its own defence.

Nvidia’s startup threats ranked