Wafer aimed for $18M and raised a $40M Series A instead

Why it matters

The $40M round finances Wafer's shift from GPU-kernel automation into full-stack inference, where optimization can become recurring infrastructure revenue.

Wafer raised a $40M Series A co-led by Marathon Management Partners and Chemistry, giving founders Emilio Andere (@gpuemi) and Steven Arellano (@gpusteve) fresh capital to automate the engineering work required to run AI models faster and more cheaply.

Arellano delivered the news in a 15-post thread on X on September 1st, writing that Wafer had planned to raise $18M and "ended up raising a $40M Series A instead." Wafer confirmed the amount and the full investor list in its funding announcement.

Wing, AMD Ventures and Outset Capital participated. Existing investors Fifty Years and Y Combinator also invested again. Wafer named Jeff Dean, Guillermo Rauch, Andy Fang, Kyle Vogt, Akshay Kothari, Matthew Prince and Scott Stephenson among the round's angel investors.

The financing came four and a half months after Wafer disclosed a $4M seed round led by Fifty Years, with Liquid 2 and Y Combinator participating. The two rounds bring Wafer's disclosed funding to $44M.

From GPU kernels to inference infrastructure

Andere and Arellano met as freshmen at the University of Chicago and later became roommates. Arellano worked on Bard infrastructure at Google and performance engineering at Two Sigma. Andere trained weather models at Argonne National Laboratory and published machine-learning security research at NeurIPS, according to Wafer's seed announcement.

They founded Wafer in 2025 around an agent that writes and optimizes GPU kernels, the low-level programs that determine how effectively AI workloads use accelerator hardware. Wafer has since expanded further up the serving stack. Y Combinator describes Wafer as an inference provider whose agents modify kernels, batching, scheduling and memory layouts for open-source models.

That expansion is the commercial bet behind the Series A. A kernel-writing tool addresses a specialized engineering task. Running optimized models as a service gives Wafer a recurring infrastructure product and puts Wafer closer to the cost and performance decisions made for every customer workload.

Wafer says its software studies traffic patterns and performance constraints before selecting and tuning the model, inference engine, kernels and hardware. The objective is continuous optimization after deployment, rather than a one-time performance pass by scarce infrastructure engineers. Wafer said the new capital will fund further automation of that loop.

The hardware layer is central to Wafer's pitch. Wafer runs workloads on Nvidia and AMD accelerators and plans to support additional architectures. That approach targets a persistent disadvantage for non-Nvidia hardware: years of developer attention and software tuning have accumulated around Nvidia's CUDA stack, leaving alternative accelerators with fewer mature optimizations.

AMD Ventures' participation gives Wafer a strategic investor with a direct interest in narrowing that software gap. In July, Wafer said it had optimized Z.ai's GLM-5.2 model on AMD's MI355X accelerators to deliver more than twice the performance per dollar of an Nvidia Blackwell deployment. The performance claim comes from Wafer's own testing.

A fast revenue ramp, annualized

The Information reported that Wafer was valued above $200M in the Series A, citing a person familiar with the financing. Andere declined to confirm the valuation to the publication.

Andere also told The Information that Wafer reached roughly $8M in annualized revenue within 12 weeks, or about $666,000 at the reported monthly pace. That figure is a run rate extrapolated from a short operating period rather than revenue collected over a full year. Andere identified Vercel and Inworld AI as customers and said Wafer operates at gross margins of roughly 50%, reflecting the expense of renting accelerator capacity.

The Information also reported that Wafer rejected multiple acquisition offers from larger inference and cloud providers, citing a person familiar with those approaches. The $40M round gives Andere and Arellano the capital to remain independent while building a broader serving platform.

Wafer is entering a capital-intensive market where raw model speed is only part of the sale. Customers also care about availability, reliability, latency under real traffic, data handling and price stability. The Series A buys Wafer time and compute to prove that autonomous performance engineering can produce durable infrastructure margins, rather than a sequence of impressive benchmark results.