Wafer

Overview

Wafer builds AI agents that tune how models run in production. For a given workload they search across the model, the serving engine, low-level GPU kernels and the hardware itself for the fastest and cheapest setup, and Wafer sells the result as hosted open-source models behind a standard API, on both NVIDIA and AMD chips. The company is going after two problems at once, the shortage of performance engineers and the software gap that keeps buyers tied to a single chip vendor.

ARR is from a third-party tracker, September 2026, and is not company-confirmed.

Thesis

NVIDIA’s position in AI computing rests on its software as much as its hardware. Since introducing CUDA in 2007 it has built up a large body of optimized code along with compilers, profilers and debuggers, and most AI frameworks assume NVIDIA hardware by default. Competing accelerators often arrive with comparable specifications and lower prices, then deliver a fraction of their expected performance because the programs that execute a model’s arithmetic were never tuned for them.

That tuning is skilled manual work. Each operation in a model runs as a small program called a kernel, and a kernel written for one chip, batch size and sequence length can be several times faster than a generic one. The serving engine above the kernels has dozens of settings for batching, memory and parallelism, and above that sit choices about numerical precision, draft models and how a model is split across chips. All of these interact with one another and with the traffic a deployment receives, and few people can work through them well. Wafer’s 2025 launch materials claimed that leading GPU engineers are paid more than $2 million a year and that most teams use less than half of the hardware they pay for.

Teams also tend to optimize once, before launch. Traffic shifts afterward, new model versions ship, new hardware arrives, and the tuning goes stale. Gartner expects inference to become the majority of AI infrastructure spending in 2026, and as it does, the cost of running below a chip’s potential grows with it.

Usage keeps raising the stakes. According to Contrary Research, agentic systems use five to 30 times more compute per task than a standard chatbot, and the AI inference market is projected to grow from $106 billion in 2025 to $255 billion by 2030. Production LLM systems were already processing tens of trillions of tokens per day by November 2025, a volume at which latency and cost decide whether an AI application can run at scale.

Founding Story

Emilio Andere (CEO) and Steven Arellano founded Wafer in 2025. They met as first-year roommates at the University of Chicago.

Andere studied mathematics. He researched transformers for weather prediction at Argonne National Laboratory, worked on adversarial machine learning at the university’s SAND Lab and was an engineer at the AI research company Elicit. Arellano studied computer science and economics, did research on large language models, and worked as an engineer at Two Sigma, at Google on infrastructure for the Bard assistant, and at Sei Labs.

The pair entered Y Combinator’s Summer 2025 batch as Herdora, which is still the company’s legal name. Their first product, pitched as Cursor for CUDA, took a customer’s PyTorch code, generated custom GPU kernels for it, and then watched the deployment in production to catch and fix inefficiencies. It was aimed at machine learning teams with growing GPU bills who could neither hire optimization specialists nor pay consultants who, according to the launch post, charge $50,000 or more per project.

The founders describe the mission in broader terms. In their words, the hardest problems are now limited by the cost of intelligence per unit of energy, and they want to maximize intelligence per watt by using AI to optimize AI infrastructure. The long-term ambition, as they put it, is intelligence “too cheap to meter.”

Product

Over its first year Wafer moved from selling an optimization tool to selling what the tool produces. As of September 2026 its main offering is hosted inference for open-source models, and the optimization system is how that inference gets made.

Serverless inference

Wafer serves open-source models through an OpenAI-compatible API and keeps no customer data. The company pitches it as the fastest open-source models for enterprises at the lowest cost per token.

Optimization agents

Wafer’s agents learn a workload’s traffic patterns and its constraints, such as a latency ceiling or a cost target, and then search for the best deployment across four layers. At the model layer they adjust numerical precision and draft models, at the engine layer they tune batching and memory, at the kernel layer they write new code, and at the hardware layer they pick between NVIDIA and AMD. Each candidate is compiled, run and measured against speed and quality targets. The ones that beat the current deployment are kept, and the search goes on. A job posting describes the machinery underneath as an agent framework that iterates on kernels, profiling infrastructure hooked into both NVIDIA’s and AMD’s performance tools, and compiler tooling that analyzes low-level GPU code.

The problem suits AI agents because the feedback is fast and exact. A language model proposes a change, the system compiles and runs it, and a measurement comes back within minutes, with a second check confirming that output quality has held. The company contrasts this with current practice, which it calls manual, service-heavy and done once before deployment. Its Series A announcement says the new money will go toward automating more of the optimization loop, so that every deployment gets the equivalent of a dedicated performance engineering team. That wording implies engineers still work alongside the agents on some current results.

Multi-hardware support

On NVIDIA, Wafer and Parasail released low-precision weights of the Kimi K2.6 model for the Blackwell generation. On AMD, Wafer has published results for several frontier open models on the MI355X accelerator.

Customer

Wafer sells to companies that consume open-source model inference at scale and care about price and latency. Its API follows the same conventions as other providers’, so customers can switch to it without changing their applications, and can switch away just as easily. The Series A angels include the chief executives of Vercel, Cloudflare and Deepgram, each of which runs a business that buys or serves inference in volume.

A second group of customers runs its own hardware and wants it to perform better, including cloud platforms and chip makers. Wafer’s published work with DigitalOcean and Parasail and the investment from AMD Ventures are early signs of this segment. For AMD especially, software that makes its accelerators competitive with NVIDIA’s for production inference removes the main obstacle to selling them.

Routing layers open another path. OpenRouter, which as of August 2026 served more than 8 million developers and routed over 100 trillion tokens per month, benchmarks more than 80 providers on price, latency and throughput and can send traffic automatically to whichever performs best. For a provider that wins on cost per token and speed, a listing there turns a benchmark lead into volume without a sales force. It works in reverse too. If a competitor matches Wafer’s price and speed, the router moves the traffic away just as fast.

Market Size

Wafer sells into inference. Gartner projected in August 2026 that inference would account for $23.3 billion of AI-optimized cloud infrastructure spending in 2026. Applying its 59% inference share to its $66.1 billion total gives roughly $39 billion for 2027, a derivation Gartner did not publish itself.

From Gartner, August 2026. The 2027 split applies Gartner’s 59% inference share.

MarketsandMarkets, the firm behind the $106 billion to $255 billion figure above, projected in October 2025 that the narrower inference platform-as-a-service market would reach $105.2 billion by 2030. Deloitte put the market for inference-optimized chips above $50 billion in 2026.

How much of this Wafer can reach depends on the share of inference that runs on open-weight models and on how much of it moves to non-NVIDIA hardware. NVIDIA held roughly 80% of data-center AI accelerators as of mid-2026, and that share will move only if the software gap closes, which is the gap Wafer works on.

Competition

Wafer competes with managed inference platforms for customers, with the companies behind open-source engines for technical leadership, and with chip vendors’ own software teams over who makes alternative hardware usable. It also partners with or builds on several of them. By capital raised it is the smallest of the group, with less than half of what either engine company raised and thirty to fifty times less than the platforms.

From company announcements and Contrary Research. Totals as of September 2026.

Fireworks AI

Fireworks AI, founded in 2022, runs a managed platform for open and enterprise-owned models. In July 2026 it raised $1.5 billion in Series D funding at a $17.5 billion valuation and reported more than $1 billion in annualized revenue and 40 trillion tokens processed per day. Its own performance engineering team handles batching, caching and speculative decoding, and it has far more scale, customers and capital than Wafer.

Together AI

Together AI, founded in 2022, offers cloud infrastructure for training and serving open models. It announced $800 million in Series C funding in July 2026, bringing its total to $1.3 billion as of August 2026. It operates at far greater scale and sells training as well as inference, and Wafer’s answer is automated optimization and its results on AMD hardware.

Baseten

Baseten, founded in 2019, provides infrastructure for deploying models in production. Its $1.5 billion Series F in June 2026 brought total funding to $2.1 billion, and it has an established enterprise customer base and sales organization. It competes with Wafer on latency, throughput and cost.

RadixArk and Inferact

RadixArk and Inferact were formed around SGLang and vLLM, the two dominant open-source inference engines, and in 2026 raised seed rounds at valuations of $400 million and $800 million respectively. Wafer builds on engines of this kind and chooses among them, so their progress feeds its product. Both also sell managed hosting, and every improvement they push into the open engines shrinks the margin by which Wafer’s tuning beats the default.

Infinity

Infinity, founded in 2025, builds an AI agent that writes the inference software stack for new AI chips, including kernels, compilers and debuggers. It raised $15 million in seed funding at a $100 million valuation in July 2026. Its technology is close to Wafer’s, but Infinity sells to chip makers and Wafer sells to inference buyers.

Modular

Modular builds a hardware-agnostic AI software stack meant to reduce dependence on CUDA. A Series C in September 2025 valued it at $1.6 billion, and Qualcomm was reported in 2026 to be in advanced talks to acquire it for about $4 billion.

AMD and NVIDIA

AMD maintains its own software stack, publishes its own benchmarks comparing the MI355X with NVIDIA’s B200, and has invested in Wafer through AMD Ventures. It wants its hardware to perform well for every customer, which argues for Wafer’s techniques becoming widely available.

NVIDIA keeps improving its own inference software and can adjust effective pricing, and either move would narrow the cost advantage of alternative hardware.

Frontier labs

Frontier labs ship general-purpose coding agents that are getting better at low-level systems work. If those agents learn to tune kernels as well as a specialized system does, much of Wafer’s differentiation goes away.

Business Model

Wafer charges for inference by usage. As of September 2026 it had disclosed neither pricing nor gross margin, and the only revenue figure in circulation is roughly $8 million of ARR reported by a third-party tracker.

The economics rest on a cost advantage with two parts. AMD accelerators cost less than NVIDIA’s, and tuning lets the cheaper hardware approach the performance of the more expensive one. In July 2026 Wafer reported that AMD’s MI355X reached about 80% of the throughput of NVIDIA’s B200 at less than half the cost on the GLM-5.2 model, which works out to a cost per token roughly 38% lower.

Illustrative. Prices and the NVIDIA cost are assumptions; Wafer’s cost applies its published 62.5% ratio.

At Wafer’s reported cost ratio, a provider could charge 20% less and still earn a higher gross margin than the incumbent. Fireworks AI’s gross margin, for reference, was about 50% as of July 2025.

The two parts of the advantage last differently. Any provider that buys AMD chips gets the hardware discount. The tuning advantage lasts only as long as Wafer’s agents find improvements faster than those improvements spread to open-source engines and to AMD’s own software.

Traction

All of these figures come from Wafer’s blog and are self-reported. The models differ greatly in size, so the throughput numbers cannot be compared with one another, and none had been independently reproduced as of September 2026.

Wafer’s named partners are DigitalOcean, with which it published joint work on AMD GPUs; Parasail, with which it released optimized weights for NVIDIA’s Blackwell generation; and TrueFoundry, whose gateway integrates Wafer’s inference service. AMD is an investor through AMD Ventures.

The Series A came five months after the seed round and was ten times its size. The founders told The Information that larger inference and cloud providers had made several acquisition offers. A Y Combinator job posting from before the Series A described a team of four, and the company’s YC profile now lists ten.

Wafer has published no revenue figure of its own. On September 2, 2026 the ARR tracker arr.club reported that the company reached about $8 million in annual recurring revenue within four months of launching its cloud, which puts the launch around May 2026 and implies roughly $2 million of ARR added per month. That number needs care. Trackers like this compile self-reported or estimated figures that nobody audits, arr.club lists Wafer’s growth rate as undetermined, and annualizing four months of rising usage-based revenue is the most flattering way to state it. Customer counts, concentration and retention are still undisclosed, and for a product sold through an interchangeable API those are the figures that show whether $8 million is a business or a moment.

Valuation

In September 2026 Wafer announced a $40 million Series A co-led by Marathon and Chemistry, with participation from Wing, AMD Ventures and Outset Capital and from existing investors Fifty Years and Y Combinator. The Information reported a valuation above $200 million, which the company has not confirmed. Against the roughly $8 million of ARR reported the same month, that is about 25 times revenue, ordinary for inference infrastructure growing this fast and demanding for a business whose customers can leave by changing a configuration. Angels in the round included Jeff Dean, who left Google in August 2026 to co-found a new company, along with Guillermo Rauch of Vercel, Matthew Prince of Cloudflare, Scott Stephenson of Deepgram, Andy Fang of DoorDash, Akshay Kothari of Notion and Kyle Vogt of The Bot Company.

The Series A followed a $4 million seed round in April 2026, led by Fifty Years with participation from Liquid2 and Y Combinator and angel checks from Jeff Dean and OpenAI co-founder Wojciech Zaremba.

One widely syndicated notice of the Series A describes Wafer as a maker of inference chips run by different people. That description is wrong, and it is a reminder of how unreliable secondary sources are on very young companies.

Key Opportunities

Demand for a second hardware supplier

Enterprises and cloud providers want an alternative to a single accelerator vendor because of pricing, supply constraints and regional availability. If Wafer can deliver comparable performance and reliability on AMD and other accelerators, buyers could use it to adopt a second supplier without building that engineering capability themselves, and AMD’s investment suggests the chip maker sees Wafer in that role. Every point of accelerator share that leaves NVIDIA needs software of the kind Wafer produces.

Continuous optimization as a product

Compilers started as a convenience that hardware makers supplied and became a permanent layer of computing, because translating intent into efficient execution has to be redone whenever hardware or software changes. Inference optimization works the same way, since it has to be redone whenever a model, a chip or a traffic pattern changes. A system that does the work continuously, and builds up data on what works for which hardware under which conditions, could become a standing part of the inference stack in place of a one-time service. If customers come to expect their deployments to improve without intervention, whichever system does that best would be hard to displace.

Licensing to infrastructure providers

Wafer’s work with DigitalOcean and Parasail points to a business in which its system tunes other providers’ infrastructure. Selling the optimizer to clouds, chip makers and large enterprises that run their own hardware would earn software margins, reduce exposure to price wars in hosted inference, and put the company inside customers’ environments, where switching is harder. Acquirers have already paid for this layer. In August 2026 Stripe reportedly agreed to acquire OpenRouter for $7.5 billion, and AMD acquired the inference chip startup Taalas the same month.

Distribution through routing layers

Inference routing platforms rank providers on price, latency and throughput and can shift traffic automatically. In those marketplaces a measurable cost and speed advantage turns into demand with little sales effort, which suits a small company with strong benchmarks.

Key Risks

Commoditization of optimizations

Every improvement Wafer’s agents find ends up as a piece of code or a configuration, and code spreads. Open-source engines absorb good techniques quickly, and AMD runs a large software organization with the same goal. If most of Wafer’s gains become free within months, its cost advantage shrinks to the hardware discount, which any provider can get. Wafer’s answer is that optimization never stops, and that a system which keeps finding new improvements for each workload is worth more than any single improvement copied into an open-source engine. Whether that holds in practice is the central question for the business.

General-purpose coding agents

Wafer’s differentiation depends on a specialized agent system beating general-purpose tools at systems optimization. Frontier coding agents are improving quickly, and the largest inference providers have the resources to use them in-house. Infinity, which applies a similar approach to the software stacks of new chips, and the performance teams at Fireworks AI, Together AI and Baseten could all adopt the same methods.

Dependence on AMD and partner conflicts

Wafer’s headline results depend on AMD’s pricing and supply, and NVIDIA could narrow the gap through pricing or software. Several partners, including DigitalOcean and Parasail, also sell inference, and the engine companies Wafer builds on sell competing hosting. AMD cuts both ways. It supplies the hardware behind Wafer’s cost advantage and has invested in the company, and it also maintains its own software stack, publishes its own benchmarks against NVIDIA and benefits commercially if the techniques Wafer develops become standard and free.

Low switching costs and unverified claims

Hosted inference for open models is sold through interchangeable APIs, so customers can leave as easily as they arrive, and no retention data is public. Wafer’s benchmarks are self-reported, and performance benchmarks are sensitive to batch size, sequence length and other conditions. If the results fail independent testing or do not carry over to customers’ own workloads, the company’s central claim falls with them.

Compressing inference prices

Per-token prices have fallen steeply. According to Contrary Research, inference costs dropped by a factor of 1,000 between 2022 and 2024 and are projected to fall another 90% by 2030. A provider that charges per token has to grow volume faster than prices fall just to grow revenue, and falling prices shrink the dollar value of any percentage cost advantage unless volumes are very large.

Summary

Wafer builds AI agents that tune inference across the model, the serving engine, GPU kernels and hardware, and sells the results as fast, low-cost hosted open-source models on NVIDIA and AMD accelerators. It started in Y Combinator’s Summer 2025 batch as a developer tool for generating GPU kernels and has since moved to selling inference directly.

As of September 2026 Wafer had raised about $44 million, including a $40 million Series A at a reported valuation above $200 million, with AMD Ventures and Jeff Dean among its backers. Its reported results show AMD hardware reaching about 80% of NVIDIA B200 throughput at under half the cost, and the whole argument rests on that result. Wafer takes cheaper silicon, makes it competitive with software, and sells it as tokens.

What decides the outcome is whether that software stays proprietary. Every optimization the agents find is code, and code spreads into vLLM, SGLang and AMD’s own stack. Wafer is betting that a system which keeps finding new improvements for each workload stays ahead of the spread of any single one. If the bet pays off, continuous optimization becomes its own layer of the inference stack. If it fails, Wafer ends up reselling AMD’s discount alongside everyone else, against competitors with forty times its capital.

The company has published no revenue of its own, no customer count and no independently reproduced benchmark, and the $8 million ARR figure attached to it comes from a third-party tracker.