Wafer Raises $40M to Make Open-Source AI Inference Enterprise-Grade
A San Francisco startup with fewer than ten employees has raised a $40 million Series A after reaching $8 million in annual recurring revenue in roughly three to four months: that is the Wafer story, announced on September 6, 2026, and the speed of it is the point. The company optimizes inference for open-source AI models across NVIDIA, AMD, and other hardware, and it does it continuously rather than through the manual, expert-heavy tuning that most teams can afford to run only once. Marathon and Chemistry co-led the round at a valuation of approximately $200 million, a 50x step-up from a $4 million seed closed just five months earlier.
For enterprise teams navigating the accelerating AI pricing war, Wafer’s growth rate is a signal worth parsing carefully. It suggests that open-source model serving has crossed a threshold where the cost and customization advantages are compelling enough to drive rapid adoption, and that the bottleneck is no longer the model itself but the infrastructure that makes running it fast and cheap.
What Wafer Actually Does
The core product is a high-performance inference cloud focused on open-source models. Enterprises connect to serverless or dedicated endpoints and get throughput and latency that Wafer claims is competitive with much smaller or proprietary alternatives, including on hardware that is not NVIDIA.
The differentiation sits in how Wafer achieves that performance. Rather than a one-time optimization process handled by human experts, the company’s system automates kernel compilation and hardware tuning on a continuous basis. As models update, hardware mixes change, and traffic patterns shift, the optimization loop runs without manual intervention. That is a fundamentally different operating model from what a typical infrastructure team can sustain internally.
An independent validation came from Vercel, which tested Wafer on GLM-5.2 and reported roughly 2x throughput compared to other serverless providers. DigitalOcean reported significant speedups on AMD hardware for models such as Kimi. Neon Health cited the lowest observed latency that holds under load.
The Numbers Behind the Round
| Metric | Value |
|---|---|
| Series A amount | $40 million |
| Post-round valuation | Approximately $200 million |
| Team size at raise | Under 10 employees |
| ARR at raise | $8 million |
| Revenue growth | 30x in 3 months |
| Peak throughput | 2 trillion tokens per month |
| Seed round (April 2026) | $4 million |
| Total funding | Approximately $44.5 million |
| Seed-to-Series A step-up | Approximately 50x in 5 months |
Wafer was founded in 2025 by Emilio Andere and Steven Arellano, both University of Chicago alumni. Andere brings an ML security research background with work published at NeurIPS and experience in weather modeling at Argonne National Lab. Arellano comes from high-performance computing and AI infrastructure roles at Google’s Bard team, Two Sigma, and Sei Labs. The two met in Y Combinator’s Summer 2025 batch, which anchors the company’s engineering discipline in a setting that typically separates revenue velocity from narrative.
Why Hardware Diversity Is the Real Story
AMD Ventures participated in the Series A, which is not a passive financial move. AMD is the largest NVIDIA alternative at inference scale, and a confirmed performance advantage on AMD hardware is a commercial claim that carries weight. If Wafer can make AMD competitive for open-source workloads, it creates a genuine second option for enterprises that want to break out of single-vendor GPU dependency.
Average GPU utilization in production is often cited around 20%. That number means enterprises are paying for five times as much compute as they are actually using, and most of the waste sits in the gap between what a model needs at peak and what operators have provisioned to handle it. Continuous optimization that adapts to real traffic patterns is one of the few levers that can close that gap without requiring manual rework every time the model or workload changes.
The open-source inference market has been building toward this moment for several years. Together AI raised $800 million to build the infrastructure layer for open-source models. Fireworks AI raised capital targeting the same enterprise segment. What Wafer adds is a narrower, deeper bet: not a marketplace of models but a continuously improving performance layer for the models enterprises have already chosen to run.
Why Enterprises Are Choosing Open-Source Inference
The shift toward open-source model serving is driven by three pressures that are not going away. Cost is the most visible: serving a proprietary frontier model through an API can cost orders of magnitude more than running a comparable open-source model on owned or rented compute. Customization is the second: fine-tuned models trained on proprietary data cannot be served through a third-party API without exposing that data. Governance is the third: regulated industries and sovereign AI mandates increasingly require data to remain within defined infrastructure boundaries.
Wafer’s customer list illustrates the breadth. Vercel is a developer platform with high-frequency, latency-sensitive inference needs. Inworld AI builds real-time AI characters. DigitalOcean serves the developer and SMB market. These are not the same use case, and the fact that a single optimization platform serves all three without degrading on any of them is the performance claim that the angel roster around this round is betting on.
The angels include Jeff Dean, who built Google’s large-scale training infrastructure, Guillermo Rauch of Vercel (already a customer), Matthew Prince of Cloudflare, Andy Fang of DoorDash, Kyle Vogt, Akshay Kothari of Notion, and Scott Stephenson of Deepgram. That group collectively spans every layer of the AI stack from compute to application, and their presence in a sub-10-person seed-stage company suggests they are betting on the team and the technical thesis rather than on a product that is already obvious.
Notably, Wafer turned down acquisition offers from multiple cloud and inference providers. Staying independent at this stage reflects a view that the inference optimization category will be worth more as a standalone business than as a feature inside a larger platform. Whether that bet holds depends on how quickly the hyperscalers build equivalent continuous optimization into their own managed services.
What the Capital Will Fund
Wafer is directing the $40 million toward three areas: automating more of the optimization loop, expanding GPU capacity to meet demand, and growing the team. Immediate hiring is focused on kernel engineers, infrastructure engineers, a chief of staff, and go-to-market roles, all in San Francisco and all requiring full-time, in-person presence.
Automating the optimization loop is the most important of the three. The commercial story depends on the system improving itself without human intervention at scale, and that requires investment in the kind of ML-driven tuning infrastructure that is expensive to build and hard to replicate quickly. If Wafer can extend its automation advantage before better-resourced competitors build comparable systems, the moat around its cost and performance claims becomes structurally defensible rather than contingent on a small team of experts working unusually hard.
For enterprise AI builders evaluating their inference strategy, Wafer’s raise is a useful calibration point. The company’s velocity shows that hardware-agnostic, continuously optimized inference for open-source models is not a theoretical advantage: it is a live product with paying customers, measurable throughput gains, and enough demand to outstrip GPU supply at eight-figure ARR. That combination makes it worth evaluating alongside managed alternatives for any workload where model customization and cost matter more than the convenience of a proprietary API.
If you are working through what an open-source inference strategy looks like for your enterprise AI stack, Enera can help you scope and build it.