Cerebras Has 98 Open Positions and a $20 Billion OpenAI Contract. The Engineers It Needs Already Work for Nvidia.

The Post-IPO Inference Pivot

Cerebras Systems filed its S-1 on April 17, 2026, and the document makes one thing clear: this is no longer a training-chip company trying to crack Nvidia's dominance. The Sunnyvale-based wafer-scale chipmaker is going public on the back of a massive inference deal, and the numbers in the filing tell the story of a business that has fundamentally reoriented itself around serving AI models to end users, not just building them.

The centerpiece is a Master Relationship Agreement with OpenAI valued at over $20 billion, covering 750 megawatts of low-latency inference capacity delivered through 2028, with an option to expand to 2 gigawatts by 2030. OpenAI also loaned Cerebras $1 billion at 6% annual interest in December 2025 to fund data center build-out, and received warrants to purchase 33.4 million shares of non-voting Class N stock. In a Wall Street Journal interview, CEO Andrew Feldman put it bluntly: "Obviously, [Nvidia] didn't want to lose the fast inference business at OpenAI, and we took that from them."

That quote maps directly to the technical bet Cerebras has been making for two years. The WSE-3 packs 44 GB of on-chip SRAM with 21 petabytes per second of memory bandwidth, roughly 7,000 times the bandwidth of an Nvidia H100's HBM3. For inference workloads, where the bottleneck is moving model weights to compute rather than raw FLOPS, that architecture delivers what Cerebras claims is up to 15x lower latency than GPU clusters. Sachin Katti, who leads compute infrastructure at OpenAI, said Cerebras adds "a dedicated low-latency inference solution to our platform" that means "faster responses, more natural interactions, and a stronger foundation to scale real-time AI to many more people."

The financials back the pivot. Cerebras generated $510 million in 2025 revenue, up 76% year over year, and reported non-GAAP net income of $237.8 million, the company's first profitable year after nearly a decade of losses. The filing also discloses an AWS partnership signed in March 2026 that enables cloud services on Cerebras chips and includes a $270 million stock purchase by Amazon. Two hyperscalers now hold direct financial stakes in Cerebras's inference roadmap.

CNBC reported the final price represents a 2.45x step-up from the Series H round, with the order book closing roughly 20x oversubscribed.

The training market still belongs to Nvidia. The inference market is where Cerebras believes wafer-scale economics win, and the OpenAI contract is the proof point. Whether that thesis holds depends on execution at a scale Cerebras has never attempted: deploying on the order of 30,000 CS-3 systems to meet the 750 MW commitment, with manufacturing yields at wafer scale that the company has so far only demonstrated in far smaller volumes.

Sunnyvale's Hidden Hiring Surge

Cerebras Systems has 24 open positions listed on Glassdoor for its Sunnyvale headquarters. The company's own careers page and its Greenhouse job board show 98 open roles across all departments, the vast majority anchored at its Sunnyvale office at 1237 E. Arques Ave. Zero G Talent's board alone has captured three Cerebras roles added in the past seven days, two inference platform engineers and a physical design engineer, suggesting the 24-role figure is already stale.

The hiring clusters in specific teams that map directly to inference infrastructure. The AI Cloud department alone lists six open roles in Sunnyvale, including a Staff Software Engineer, Inference Cloud, a Principal Engineer, AI Inference Reliability, and a Staff Inference ML Runtime Engineer. The Software department adds another dozen: kernel engineers, ML systems performance engineers, network architects, and an Engineering Manager, Inference ML Runtime. The Advanced Technology group is hiring compiler engineers and AI/ML research scientists across Sunnyvale, Toronto, and Vancouver.

These aren't generic software titles. The Staff Software Engineer, Inference Platform posting on LinkedIn calls for someone to "own the orchestration layer that runs inference on our datacenter clusters"—the glue between cloud and ML components. The job requires Kubernetes expertise, experience with TTFT (time-to-first-token) and tail-latency reduction, and proficiency in Go or C++. The parallel listing on Built In for a non-staff Software Engineer, Inference Platform drops the experience requirement to three years but keeps the same technical bar: distributed systems, Kubernetes CRDs, high-QPS optimization, mTLS security.

Cerebras is hiring at every level, from Member of Technical Staff up through Principal Engineer, which signals it's building out an entire org layer, not backfilling a few seats. The presence of that ML Runtime management role means the team has reached the headcount threshold where it needs dedicated leadership. That's a department, not a project.

Meanwhile, the Silicon and Systems divisions are still hiring aggressively in Sunnyvale, physical design engineers, RTL leads, post-silicon bring-up staff, mechanical and electrical engineers, which means Cerebras isn't outsourcing its hardware pipeline to support a software pivot. It's scaling both tracks at once.

Sunnyvale's AI-chip job market is more crowded than most people realize. LinkedIn's "similar jobs" sidebar for Cerebras' inference platform role surfaces competing openings at NVIDIA (Distinguished Engineer, Dynamo; Principal Software Engineer, AI Inference, both in Santa Clara), AMD (Principal AI Inference Systems Engineer), CoreWeave (Staff Software Engineer, Inference, Sunnyvale), Google (Staff Software Engineer, TPU Machine Learning Supercomputer, Sunnyvale), and Anthropic (Staff+ Software Engineer, Inference Runtime, San Francisco Bay Area). Cerebras is fishing in the same pond as its largest competitors, and it's doing so from a city that doesn't carry the talent-market reputation of San Francisco or Palo Alto. That's the hidden part: the war is being fought on Arques Avenue, not Sand Hill Road.

What the Roles Reveal About the AI Inference Stack

Reading Cerebras' job board is like looking at a blueprint of the company's inference stack, layer by layer, role by role. The company has 98 open positions on its Greenhouse board. More than half are in software departments, and the titles tell a specific story about what Cerebras is actually building: not just a chip, but a full-stack inference platform that competes with the cloud services running on NVIDIA hardware.

Start with the AI Cloud department. It holds listings for those same Sunnyvale- and Toronto-based roles: the Staff Software Engineer, Inference Cloud; a Principal Engineer, AI Inference Reliability; and a Staff Inference ML Runtime Engineer. These aren't generic cloud roles. The inference ML runtime position alone signals that Cerebras is building custom software to execute model predictions on its wafer-scale chips, the layer that translates a trained model into actual outputs at speed. The reliability role tells you the company knows that speed means nothing if the service drops under load.

Then there's the Software department, which reads like a systems-programming syllabus. Kernel Engineers, Kernel Reliability Engineers, a Staff Kernel Optimization Engineer, and an ML Systems Performance Engineer, all roles that operate at the boundary between hardware and software. Kernel engineers write the low-level code that lets an operating system talk to silicon. Cerebras needs them because its wafer-scale engine doesn't behave like a GPU; the standard software stack doesn't fit. Someone has to write the driver layer that lets PyTorch or a customer's application actually run on a chip with 4 trillion transistors.

The compiler team, listed under Advanced Technology as a Compiler Engineer and an R&D Engineer for AI/ML and HPC, fills another gap. A compiler maps high-level model code down to hardware instructions. Cerebras' architecture is novel enough that off-the-shelf compilers don't work. These engineers are building the translation layer that lets a data scientist's Python script run on Cerebras hardware without rewriting everything from scratch.

What ties these roles together is the job description for the Senior Software Engineer, AI Inference Platform, listed on BuiltIn and Employbl. The posting calls for someone to build "core APIs for the Inference Platform, handling model catalog management, deployment of ML workloads, scaling, and status monitoring." It asks for RESTful APIs, gRPC, Kubernetes, Docker, Postgres, Redis, and observability tools like Prometheus and Grafana. This is the control plane, the software that lets a customer upload a model, spin up a serving instance, and monitor performance without ever touching the bare metal.

Put it all together and the inference stack looks like this: custom kernels and compilers at the bottom, a runtime layer in the middle that executes models on wafer-scale hardware, and a platform layer on top that exposes all of it through APIs a software engineer can actually use. That's not a chip company hiring a few software people. That's a platform company that happens to make its own silicon.

The competitive signal is clear. NVIDIA's dominance in inference runs on CUDA, a software ecosystem two decades in the making. Cerebras is trying to build its own equivalent from scratch, and the 98 open roles show how much engineering that takes. Every kernel engineer, compiler developer, and platform API builder is a bet that owning the full stack, from transistor to REST endpoint, matters more than riding on someone else's software.

Why Inference, Not Training, Is the Next Battleground—and the Talent War No One's Watching

Cerebras' Sunnyvale hiring spree isn't happening in a vacuum. It's a bet on where the AI hardware market is heading, and the direction is away from training and toward inference.

For the better part of a decade, the chip industry's center of gravity was training: building ever-larger GPUs and custom accelerators to handle the massive parallel computation required to teach models like GPT-4 and Claude. NVIDIA built a $3.5 trillion company largely on that premise. But the economics are shifting. Once a model is trained, every time a user sends a prompt, the system has to run inference, generating a response, classifying an image, translating text. That happens billions of times a day across millions of users. The compute cost of serving those requests is starting to dwarf the one-time cost of training the model itself.

This is the transition Cerebras is positioning for. The company made its name with the Wafer Scale Engine, a chip designed to accelerate training workloads at a scale no competitor could match on a single die. But training is a batch process: you run it, you're done. Inference is continuous, latency-sensitive, and unforgiving. A user waiting two seconds for a chatbot response will leave. The hardware requirements are fundamentally different: lower latency, higher throughput per watt, and the ability to handle many small requests simultaneously rather than one enormous one.

The job postings reflect this. Cerebras isn't looking for training-cluster architects. It's hiring inference platform engineers and ML systems performance engineers, roles optimized for the serving side of the stack. The inference platform positions in Sunnyvale specifically call for experience with real-time deployment, model serving frameworks, and performance optimization under production load. That's a different skill set than what the company recruited for even two years ago.

Competitors see the same shift. NVIDIA's product roadmap now dedicates significant silicon and software investment to inference, with the Blackwell architecture explicitly designed to handle serving workloads efficiently. AMD has pushed its MI400 series into inference-optimized configurations. Startups like Groq, SambaNova, and Tenstorrent have raised hundreds of millions on the premise that inference is the next multi-billion-dollar hardware market. Even cloud providers, AWS, Google, and Microsoft, are designing custom inference chips (Trainium, TPU, Maia) because off-the-shelf GPUs are too power-hungry and expensive for high-volume serving.

The market logic is straightforward. Training demand is concentrated: a handful of well-funded labs and hyperscalers buy the hardware, and they buy in bursts tied to new model releases. Inference demand is distributed and growing: every company deploying an AI product needs inference capacity, and that need scales linearly with user adoption.

Cerebras' post-IPO timing matters here. Going public gave the company capital and visibility, but it also created pressure to show a path to revenue. Training chip sales are lumpy, big deals with a few customers. Inference is a recurring, usage-based market. If Cerebras can build a platform that makes its wafer-scale architecture competitive for serving workloads, it opens a revenue stream that looks very different from its training-era business model.

The Sunnyvale hiring surge is the visible edge of that strategy, and it's pulling from a shallow pool. The company's careers page shows 98 open positions across inference platform engineering, physical design, silicon, systems, and cloud infrastructure. Multiple roles list Sunnyvale as the base. Cerebras is competing for the same engineers that NVIDIA, AMD, CoreWeave, Google, and Anthropic are all recruiting in the South Bay.

The South Bay's AI-chip labor market is a closed loop: engineers at NVIDIA's Santa Clara campus, AMD's Sunnyvale office, and the various inference startups along Mathilda Avenue are separated by a short drive and a shorter LinkedIn network. When a $56-billion-valued company opens a hiring pipeline for inference-specific roles, kernel optimization, compiler design, distributed systems for wafer-scale hardware, it targets people who already work on those exact problems at the incumbents.

The startups face particular pressure. Groq, SambaNova, and Tenstorrent are all chasing non-GPU inference, and all are smaller than Cerebras. When a well-funded competitor with public-market visibility offers compensation packages that reflect a $56 billion valuation, the calculus for a senior engineer at an earlier-stage startup shifts. Cerebras doesn't need to outbid everyone. It just needs to be the most credible alternative to NVIDIA, and at its current valuation, it is.

The net effect is a slow drain that doesn't show up in press releases. The job postings keep multiplying, the LinkedIn profiles keep updating, and the Sunnyvale inference cluster keeps thickening. The war is happening. The casualties just haven't been named.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at Cerebras, and the people building the field.