Cerebras Grew 94%—and Margin Got Harder

Cerebras grew Q1 revenue 94% and narrowed its net loss 41%, then warned that core gross margin would shrink. The warning came as AWS planned to deploy its wafer-scale engine beside cheaper Trainium processors. The wafer remains remarkable. The unresolved question is who owns the route from that wafer to a paid token.

The proof moved from silicon to margin

In 2019, Cerebras answered a physical question with a physical object. Its CS-1 system placed the world’s largest chip at Argonne National Laboratory, where the first machines were used for basic research. By 2021, its WSE-2 carried 850,000 AI-optimized cores and 2.6 trillion transistors. The escalating numbers documented technical progress and made wafer-scale computing legible against an industry organized around smaller chips.

Then the question changed. Cerebras reported Q1 revenue of $193.4 million and a $14 million net loss. Those are not signs of absent demand or a failed product. But the company also forecast that its core gross margin would shrink in Q2, moving the test from whether customers would deploy wafer-scale computing to whether those deployments could generate durable operating leverage.

Cerebras’ May IPO made that change formal. The company raised $5.55 billion at a $56.4 billion fully diluted valuation, and its shares closed their first trading day up 68%. The first-day jump validated investor appetite, but public shareholders replaced technical possibility with quarterly evidence. A research customer could ask whether the machine worked. A shareholder has to ask what remains after the machine is delivered.

Inference growth intensifies the price war

Specialists appear to benefit as inference expands the compute required after a model has been trained. Barclays projected that inference capital spending would surpass training spending within two years, creating a large market for systems designed around fast model serving rather than occasional training runs.

But the supplier’s opportunity is also the buyer’s expense. OpenAI and Anthropic told investors that inference costs exceeded half of their revenue. Model providers do not buy accelerators to protect accelerator pricing; they buy them to reduce the cost of producing outputs. They can use every improvement in hardware, software or model execution to renegotiate the serving layer.

OpenAI engineers reportedly told colleagues that they had found a way to more than halve inference costs. The competitive baseline can move without waiting for a new chip generation. Chipmakers can create value in silicon, but a software change can reduce how much of that value the model provider must surrender to the supplier.

The specialist’s performance can be real without being the customer’s objective. Model providers want the lowest cost for an acceptable output, repeated across every request. As inference consumes more of their revenue, they have stronger incentives to capture savings internally. Technical gains that once supported premium pricing become inputs to the buyer’s margin plan.

AWS turned the breakthrough into a service tier

AWS made the new competitive unit explicit. The cloud provider plans to deploy Cerebras’ Wafer-Scale Engine for fast inference while continuing to offer slower, cheaper computing through its own Trainium processors. Cerebras earned a place inside one of the largest cloud portfolios.

But AWS controls the menu on which Cerebras appears. It places the specialist beside a lower-cost internal alternative and can segment workloads according to what customers will pay for speed. Cerebras supplies differentiated performance; AWS owns the distribution, customer relationship and adjacent substitute.

AWS is building heterogeneous AI compute as an economic system, not merely a hardware inventory. Different accelerators do not just coexist in a data center. AWS can choose which machine runs, which price reaches the customer and how much utilization each supplier receives.

Performance can win a lane without winning the stack. Cerebras may be the appropriate engine when latency carries a premium, while Trainium remains available when lower cost matters more than speed. AWS does not need one architecture to dominate every request. It needs enough architectures to direct each request toward the economics it prefers.

Frontier model operators are building the same optionality from the other direction. OpenAI and Broadcom developed Jalapeño, an inference chip optimized for large language models, from design through manufacturing tape-out in nine months, with OpenAI’s models assisting the process. Meta had already put both MTIA v1 and its next-generation training-and-inference accelerators into production. These companies can buy external systems, develop internal chips and change software execution at the same time.

AWS does not need to eliminate Cerebras to gain leverage. Deploying it beside Trainium gives AWS speed at the premium tier and an internal option when customers prioritize cost.

The rack now carries a capital structure

A benchmark stops at the accelerator. A paid token comes from a powered rack in a completed data center, connected to capacity reserved before the request arrived. The economic unit has widened from chip performance to site execution, utilization, power and financing.

Morgan Stanley estimated that $2.9 trillion of AI infrastructure spending through 2028 would be funded roughly $1.4 trillion by hyperscalers and $1.5 trillion by debt, private equity, venture capital and other sources. By May 2026, hyperscaler unsecured-bond issuance had reached $155 billion for the year, more than 45% above all of 2025. Some AI-infrastructure bond sales were four times oversubscribed.

Hyperscalers had announced 46 gigawatts of AI data-center capacity, enough at full utilization to consume as much energy as roughly 44.2 million US households. At that scale, accelerator performance is one variable inside a system constrained by substations, sites, capital and contracts.

When customers make long-duration capacity commitments, their promised demand helps finance data centers. Completed sites give integrated providers capacity to offer lower prices or bundled terms, drawing more workloads into the same stack. The loop begins with the ability to turn future demand into present concrete, power equipment and racks.

Cerebras’ IPO gives it billions of dollars to participate in this buildout, countering any simple story that standalone vendors are being erased. Yet Cerebras competes with hyperscaler balance sheets, bond markets and infrastructure investors financing capacity through 2028. Benchmark performance cannot lower the cost of that capital, secure a powered site or keep an underused rack economically whole.

Independence survives, but not on benchmark terms

Integrated buyers still need differentiated technology, supply alternatives and talent, so specialists have not disappeared. But specialists increasingly reach demand through larger institutions. Cerebras uses AWS for cloud distribution. SambaNova raised $1 billion at an $11 billion valuation and signed JPMorgan to deploy its chips for in-house AI. Nvidia and Groq reached a $20 billion non-exclusive inference licensing agreement, while several Groq senior executives were set to join Nvidia.

Cerebras gains volume through AWS but sits beside Trainium. SambaNova’s JPMorgan anchor validates deployment but gives one customer outsized influence over utilization. Groq’s license monetizes intellectual property while Nvidia captures more downstream economics. Arm and SoftBank had also shown preliminary acquisition interest in Cerebras before the IPO, confirming the strategic value of its silicon while the economics of independent scale remained unsettled.

Cerebras rejected that interest, raised public capital and entered this phase as a live participant, not a stranded invention. It can now fund deployments without selling itself. But an enterprise buyer choosing inference capacity will compare a Cerebras-powered premium tier with Trainium, internal silicon and software cost cuts—not with a benchmark in isolation.

In 2019, Cerebras could place its answer inside a single wafer and send the machine to a research laboratory. By 2026, AWS could place that wafer beside Trainium and sell its advantage as one tier among several. That is how revenue can rise 94% while management forecasts a lower core margin: the wafer supplies the speed, but the stack decides who keeps the gain.