Signals Inbox·July 20, 2026·AI Chips

Groq vs Cerebras: who is winning?

Cerebras is winning the race to become the strongest independent high-speed AI inference platform, while Groq still leads in developer reach and global self-service access.

Q1Why are Groq and Cerebras always compared?

Groq and Cerebras are compared because both companies built specialized AI processors to make model responses much faster than conventional GPU infrastructure. They sell more than chips: each company combines custom silicon, software, servers, data-center capacity and a cloud API aimed at real-time inference.

The overlap became obvious once developers could run many of the same open models through either provider. GroqCloud and Cerebras Inference compete for workloads such as GPT-OSS, Llama, Qwen and other open-weight models where users care about output speed, time to first token, price, model availability and production reliability.

Their hardware could hardly look more different. Groq connects many compact Language Processing Units and schedules every operation in advance, producing highly predictable latency. Cerebras turns an entire silicon wafer into one processor, with 900,000 compute cores, 4 trillion transistors and enormous on-chip memory bandwidth. Both designs respond to the same problem: modern models often spend too much time moving data and too little time doing useful computation.

The rivalry has become sharper lately because both companies are moving beyond impressive demos. Groq is building a distributed inference cloud and fitting it with next-generation systems. Cerebras is expanding its cloud, serving frontier-model workloads, entering a major cloud platform, adding European capacity and increasing CS-3 manufacturing. The winner will be the company that can turn unusual hardware into dependable infrastructure at industrial scale.

Q2Why is Groq vs Cerebras still so hard to call?

Groq and Cerebras look stronger on different kinds of evidence, so a quick verdict misses the real tension. Groq publishes adoption and infrastructure figures. Cerebras, as a public company, publishes revenue, margins, customer concentration, losses, cash, contracts and future obligations.

Groq’s disclosures show reach: millions of developers, thousands of AI-native companies, trillions of weekly tokens and a presence across four major regions. Those numbers tell us that engineers are trying the product and that meaningful traffic is running through GroqCloud. They do not tell us how many accounts pay, how much revenue each token produces or whether usage is concentrated in heavily subsidized partnerships.

Cerebras gives us the opposite picture. Its filings show exactly how much money the company earns and how quickly cloud revenue is growing, but the customer base remains concentrated and much of its promised future scale still has to be built. A huge contract can be more valuable than millions of registrations. It can also become a serious execution problem when one customer dominates the buildout.

That leaves several questions to settle: who has real revenue, who is growing into paid demand, who is faster, who can add capacity and who still controls the technology.

Q3Is Cerebras already a much bigger business than Groq?

The only financial numbers we can verify put Cerebras far ahead of Groq today. Cerebras reported $193.4 million of quarterly GAAP revenue, up from $99.5 million one year earlier. Its latest full-year outlook calls for $855 million to $865 million of core revenue, roughly 69% growth at the midpoint.

Cerebras generated $110.6 million from hardware and $82.8 million from cloud and other services in the quarter. Cloud already represented about 43% of total revenue, so the company is no longer growing mainly by shipping occasional supercomputers. A large and fast-growing part of the business now comes from the same hosted inference workloads that Groq wants to dominate.

Groq remains private and does not report audited revenue. Previous press reports discussed internal targets and estimates, but placing an unverified projection beside a public filing would make the comparison look cleaner than it is. Cerebras has proved it can make hundreds of millions of dollars. Groq has proved it can attract a large audience.

Commercial position of Cerebras and Groq

Q4Is Cerebras growing faster than Groq right now?

Disclosed revenue is growing faster at Cerebras, while developer registrations are growing faster at Groq. We give revenue more weight in the final verdict, but Groq’s developer expansion is too large and too consistent to dismiss.

Cerebras nearly doubled quarterly revenue, with GAAP sales rising 94% year over year. Cloud and other services grew 178%, far faster than the 59% increase in hardware. Hosted inference is now growing much faster than hardware and becoming a larger part of Cerebras.

Groq’s developer count rose from more than two million in September 2025 to 3.5 million in February 2026 and more than five million four months later. That works out to at least 150% growth in roughly nine months. The progression across three separate disclosures makes it more credible than a one-off registration campaign. Groq also says production traffic has kept rising, although it has not published a precise series that outsiders can audit.

The missing piece is conversion. Cerebras has shown that growth turns into money; Groq has shown that more people keep trying its service. What we need next is paid token growth or revenue per active customer, not another larger registration total.

Q5Does Groq’s huge developer audience prove that GroqCloud is winning?

GroqCloud clearly attracts more developers than Cerebras today, but we still cannot tell whether Groq makes more money from them. Developers can start quickly, use an OpenAI-compatible API, access a generous free tier and run popular models at unusually low latency. That easy start helps Groq become the first infrastructure provider an engineer tries.

Groq’s customer and product activity makes that audience more credible. The company has powered Meta’s official Llama API, become the exclusive inference provider for Bell Canada’s sovereign AI network, supported HUMAIN’s enterprise AI platform and added production-oriented features such as prompt caching, remote MCP support, private tenancy and compound AI systems. Groq is trying to move users from experiments into workflows that are harder to replace.

What we still cannot see is how many of those developers pay. Groq does not reveal paid developers, annual contract value, renewal rates, net revenue retention, utilization by data center or the share of traffic coming from its free tier. A large developer base can help adoption long before it produces much revenue, and Groq may already be converting users better than outsiders realize. The company has simply not shown it.

Q6Is Cerebras’s $25 billion backlog as strong as it sounds?

Nothing Groq has disclosed matches Cerebras’s $25 billion of remaining performance obligations, although the headline overstates how quickly the money will arrive. Most of the balance comes from the OpenAI agreement, under which Cerebras must provide 750 MW of inference capacity over several years.

The contract is unusually concrete. OpenAI is committed to purchasing the initial capacity, supplied Cerebras with an approximately $1 billion working-capital loan and received an option for another 1.25 GW by 2030. Cerebras says the committed agreement is worth more than $20 billion. Customer financing, contracted capacity and optional expansion go far beyond the standard partnership press release.

The accounting schedule is slower than the headline suggests. Cerebras expects to recognize about 16% of the obligations during the first 24 months, another 45% during months 25 to 48 and the rest later. Part of the amount covers data-center costs that Cerebras passes through to OpenAI, so not every dollar carries attractive technology margins.

The latest quarter shows that the OpenAI ramp has barely started. The contract contributed $16.9 million of revenue, less than 9% of the total. It has already transformed Cerebras’s financing and capacity plans, but barely touched the income statement.

Q7Is Cerebras inference actually faster than Groq inference today?

Cerebras currently wins the clearest shared large-model benchmark by a wide margin, while Groq ranks first on the current Llama 3.3 70B table.

Artificial Analysis currently measures GPT-OSS-120B at 1,745 output tokens per second on Cerebras in the high-reasoning setting, versus 478 on Groq. Cerebras also returns the first answer token in 1.65 seconds, compared with 4.90 seconds for Groq. The low-reasoning setting produces almost the same pattern: 1,803 tokens per second for Cerebras and 477 for Groq. Cerebras is about 3.7 to 3.8 times faster on the cleanest live comparison we can make.

Older comparisons point in the same direction, although they should not be mixed mechanically with today’s endpoints. During 2025, same-model tests on GPT-OSS-120B, Llama 4 Maverick and Llama 3.3 70B showed gaps closer to five or six times. Meanwhile, the current Artificial Analysis page for Llama 3.3 70B no longer lists Cerebras and ranks Groq first at roughly 311 tokens per second. Any claim that Cerebras is faster needs a model name attached to it.

Cerebras gets its peak speed by keeping an exceptional amount of compute and memory bandwidth on one wafer, cutting communication delays between chips. Groq distributes a statically scheduled program across many smaller LPUs. Groq’s approach offers predictable latency and modular deployment, but it has a lower maximum output rate on the largest shared benchmark available now.

Speed alone cannot settle every workload decision. Input length, concurrency, numerical precision, price, availability, model choice and answer quality all affect the result. A 2026 academic comparison also found that the best accelerator changed with model size, sequence length and batch size, while several specialized systems consumed more power when idle. Cerebras wins the peak-speed comparison today. Customers still need enough utilization to make the system economical.

Current independent inference comparison

Q8Is Cerebras solving the real AI inference bottleneck better than Groq?

For raw large-model delay, Cerebras is ahead. For easy access from many locations, Groq is ahead.

Cerebras attacks the delay inside the machine. Its wafer-scale processor cuts the time spent moving model data between chips, which is why it can produce long outputs at speeds that conventional clusters and Groq often do not match. OpenAI’s use of Cerebras for an interactive coding model shows where that performance becomes commercially useful: users can watch code appear almost immediately instead of waiting through a long generation.

Groq attacks the delay between the customer and the machine. Its cloud is spread across several regions, its API is familiar and its smaller units can be deployed more incrementally. Those choices help Groq support sovereign installations and put inference closer to users without requiring every customer to adopt a giant dedicated system.

Capacity is the real test for Cerebras. The company recently announced 200 MW of European compute by the end of 2027 and new Flex manufacturing lines intended to support a sevenfold increase in CS-3 production. Those announcements address the obvious concern that a spectacular processor may be difficult to build and deploy in sufficient volume. Now the factories and data centers have to deliver.

Groq also targets 200 MW by the end of 2027, but across its global cloud. The identical headline number hides a striking difference: Cerebras expects its European footprint alone to match Groq’s worldwide target. Should both companies deliver, the capacity race moves sharply toward Cerebras.

Q9Does Cerebras have better customers than Groq?

Cerebras has the most valuable customer contract and the stronger cloud-platform partnership, while Groq appears to have the broader mix of production relationships. OpenAI can fill a huge amount of capacity, and AWS can put Cerebras inside existing cloud accounts. Groq’s work with Bell, Meta, HUMAIN, Paytm and regional infrastructure partners spreads its bets across more customers and countries.

OpenAI is the standout relationship. It has committed future spending, lent money for the buildout, begun using Cerebras for Codex-Spark and kept an option to buy much more capacity. The AWS relationship also goes well beyond logo-sharing: the planned service places Cerebras systems inside AWS data centers, combines Trainium for prefill with CS-3 for decode and exposes the result through Amazon Bedrock.

Groq’s customers are spread across more sectors and countries. Bell chose Groq exclusively for its Canadian sovereign inference network. Meta used Groq to accelerate the official Llama API. HUMAIN selected Groq for enterprise and national AI infrastructure in Saudi Arabia. Paytm brought Groq into a large payments platform in India. None of those disclosed relationships matches the potential value of OpenAI, but together they show that Groq can sell across several industries and geographies.

Cerebras’s customer depth comes with extreme concentration today. Its latest filing shows that MBZUAI produced 63% of quarterly revenue and G42 another 11%. The two are related parties, giving one connected ecosystem 74% of sales. Three customers also represented 86% of accounts receivable. The new anchor customer should diversify revenue if deployment ramps as planned, but Cerebras remains unusually concentrated.

Groq may have its own concentration problem, especially around large sovereign projects. Private-company reporting means we cannot measure it.

Q10Did Groq win or lose its race with Cerebras by licensing technology to Nvidia?

The Nvidia deal proved Groq’s technology belongs in the top tier, but it also handed Nvidia much of the best distribution opportunity. Nvidia’s decision to license Groq technology, hire founder Jonathan Ross, president Sunny Madra and other team members, then build Groq 3 LPX into Vera Rubin confirms that the LPU architecture solved a problem Nvidia considered important.

Nvidia’s LPX rack links 256 Groq 3 LPUs, with 128 GB of SRAM and immense memory bandwidth, to accelerate token generation alongside Rubin GPUs. Nvidia says the combined platform can deliver far more inference throughput per megawatt on trillion-parameter models. Those projections still need production evidence, but this has moved well beyond a research partnership. LPX is part of Nvidia’s commercial platform.

Groq remained independent after the licensing agreement. By its latest funding announcement, Adam Winter had replaced Simon Edwards as CEO, and Groq had recruited a new operations and product team while raising $650 million. Groq now plans to install new inference systems, including Nvidia LPX, across its own footprint. It gains access to Nvidia’s hardware pipeline and can focus on operating the cloud rather than funding every layer alone.

What Groq gave up was control. Nvidia owns the strongest global distribution path for technology derived from Groq, and several of the people most closely associated with the architecture moved with it. Groq can still build a valuable cloud business, but its long-term advantage will depend more on how well it runs data centers, serves developers, keeps customers and fills capacity.

Cerebras still wants to do everything itself: design the processor, manufacture the systems, operate the cloud, sell capacity and keep the architecture under one corporate roof. The plan is harder. Cerebras keeps the upside if it works.

Q11What can Cerebras do that Groq cannot easily copy?

Wafer-scale computing gives Cerebras the harder advantage to copy, while Groq offers the easier service to start using. Wafer-scale computing requires unusual chip design, defect tolerance, packaging, cooling, compiler work, system engineering and years of manufacturing experience. A competitor cannot recreate that combination by designing one faster accelerator.

Cerebras has also learned how to route around manufacturing defects, connect the wafer to external memory, link systems for larger models and run production workloads through a specialized software stack. The new Flex lines add another layer: higher volume can improve manufacturing knowledge and spread fixed engineering costs across more systems.

Groq’s advantages come from the service around the chip. Deterministic execution produces consistent latency, GroqCloud has accumulated a large developer audience, and the company has built regional infrastructure and enterprise integrations. Applications become harder to move once they depend on Groq’s response times, APIs, reserved capacity and operational support.

Groq’s technology advantage is less exclusive than it once was, so the company increasingly has to defend itself through cloud operations, regional capacity, software and customer relationships. A much larger platform can reproduce parts of those advantages more easily than it can reproduce wafer-scale manufacturing.

Q12Can Cerebras make good money from all this growth?

Cerebras improved its margins in the latest quarter, but the capacity build is set to drag them down sharply. The latest quarter produced a 44.6% GAAP gross margin, up from 41.8% one year earlier, and the operating loss narrowed from $28.5 million to $15 million even as revenue nearly doubled.

Cerebras also made better margins on cloud services. Higher pricing and better capacity utilization helped cloud and other services reach a 48.9% GAAP gross margin. The hosted product can make reasonable money once the machines are busy.

The next quarter will look rougher. Cerebras expects core gross margin to fall to 36% to 38% and core operating margin to reach negative 30% to negative 32%. The company has also signed $2.3 billion of future lease payments for data centers that had not yet commenced at the end of the quarter. Building capacity ahead of revenue creates a period of heavy depreciation, rent, power, staffing and financing costs.

Some contract-related revenue will simply reimburse data-center expenses and add very little gross profit. Reported sales may therefore grow faster than the company’s actual profit potential. Core gross profit, utilization and cash generation will tell us more than the largest possible revenue number.

Groq offers no comparable margin or burn disclosure. A distributed cloud may use smaller units more flexibly, but a wide footprint can also carry expensive unused capacity. Groq’s economics remain an open question rather than an advantage.

Q13Does Cerebras have more money to scale than Groq?

The money gap favors Cerebras by a wide margin today, although Cerebras also has the larger bill to pay. The company raised $1 billion before its IPO, received the customer working-capital loan, collected $6.4 billion of gross IPO proceeds and arranged a revolving credit facility that expanded to $850 million. Before the IPO proceeds arrived, Cerebras already held about $3.3 billion across cash, restricted cash and short-term investments.

Groq is still exceptionally well funded. It raised $750 million at a $6.9 billion post-money valuation in September 2025 and another $650 million in June 2026. The two rounds supplied $1.4 billion in fresh capital within nine months, enough to keep expanding GroqCloud and fit out existing locations toward its 200 MW goal.

The spending plans explain why Cerebras needs so much money. Cerebras is financing wafer-scale systems, new factories, European capacity, major customer infrastructure, cloud deployments and billions in future leases. Groq is leaning more heavily on hardware developed with Nvidia, which may reduce its own chip-development burden but still leaves substantial data-center spending.

Cerebras has public-market access, strategic customer financing, bank credit and several billion dollars of equity capital. The advantage holds only if management avoids locking that money into underused capacity or weak-margin contracts.

Q14Are Groq and Cerebras actually threatening Nvidia?

Nvidia is still far ahead of both Groq and Cerebras. It has the broader hardware portfolio, the deepest software ecosystem, stronger networking, larger supply commitments and customer relationships that span both model training and inference.

Even so, both startups have already pushed the market toward mixed systems. Cerebras persuaded Amazon to pair Trainium with CS-3 instead of forcing one chip family to handle every stage of inference. AWS’s choice shows that specialized decode hardware can earn a place inside a much larger platform.

Groq’s influence is even more direct. Nvidia placed Groq 3 LPX beside Vera CPUs, Rubin GPUs, networking and storage in its flagship system. Groq’s technology may reach more customers through Nvidia than Groq could have reached alone.

That outcome makes the market harder for Cerebras. It no longer competes only against conventional GPUs; it will increasingly compete against an Nvidia platform that includes inference-specialized LPUs. Cerebras still has a major speed advantage on several shared models, but Nvidia can combine fast-enough performance with easier procurement, familiar software and a much wider workload range.

The market is likely to stay mixed. GPUs and general accelerators will handle broad workloads, while LPUs, wafer-scale systems, TPUs, Trainium and other specialized processors take the jobs where latency or economics justify switching. Groq has already shaped Nvidia’s answer. Cerebras is building the strongest independent alternative.

Q15Who is winning right now, Groq or Cerebras?

Cerebras is winning right now, and the lead is meaningful. Groq leads developer adoption and the current data-center footprint. Cerebras leads the independent company race because it has stronger verified revenue, faster large-model inference, deeper contracted demand, more capital, expanding manufacturing and full control of its core architecture.

Revenue and signed demand carry the most weight in our verdict. Cerebras has reported $193.4 million of quarterly revenue and guided toward roughly $860 million for the year, while Groq still withholds comparable financials. Cerebras also has a binding multi-year capacity contract and a planned route into AWS Bedrock. Groq’s large partnerships come with far less transparent spending commitments, while Cerebras keeps the chip, systems, cloud and customer contracts under its own control.

Groq can still close the gap. Its developer growth is unusually fast, its cloud is already global and the product is still fast enough to matter. Groq could narrow the gap quickly by disclosing substantial revenue, demonstrating high paid utilization across its sites, turning sovereign partnerships into repeatable contracts and showing that the post-Nvidia company can still create product advantages of its own.

Cerebras has more to execute. It must deliver the contracted capacity on schedule, get the AWS service live and widely used, build the announced European footprint, complete the sevenfold manufacturing expansion, diversify away from the UAE customer cluster and recover margins after the current investment phase. Missing several of those milestones would make the backlog look less valuable and Groq’s smaller buildout more attractive.

Groq is winning developer attention. Cerebras is winning the more valuable race to become a large independent inference platform.

Groq vs Cerebras: the final scorecard