# Wafer Aiolum story: https://aiolum.com/story/wafer/ · added 2026-11-06 · updated 2026-11-06 **In short.** Wafer (AI agents that tune inference for open AI models, sold as an inference API, United States): US$8M annual revenue reported, September 2026. Evidence grade D: reported, not verified. Team: 2 co-founders (Y Combinator, Summer 2025); 5 people signed the April 2026 seed post; 8 employees at the September 2026 figure (Lean AI Leaderboard); 18 people on the team page and 15 on YC's page, November 2026. Launched July 2025; still running. AI agents that tune inference for open AI models, sold as an inference API · United States · AI-native · started small, grew after funding · still running · group: Developer tools and coding agents - Outcome: US$8M annual revenue reported, September 2026 (evidence grade of the headline number: D; headline in US dollars: 8000000) - Launch: July 2025; months from launch to the outcome: 14 - Team: 2 co-founders (Y Combinator, Summer 2025); 5 people signed the April 2026 seed post; 8 employees at the September 2026 figure (Lean AI Leaderboard); 18 people on the team page and 15 on YC's page, November 2026 - Outside funding: US$44M in total: US$4M seed led by Fifty Years (April 2026) and US$40M Series A co-led by Marathon and Chemistry (September 2026), as the company states - Main factor: Performance engineering - Meters, 1 to 4: evidence strength 1, copyable pattern 1, capital needed 3 - Tags: AI-native, Small team, Funded, Running, B2B, AI infrastructure Links: - Visit wafer.ai: https://www.wafer.ai/ **The product.** An inference service for open-weight AI models such as GLM, Kimi, DeepSeek and Qwen. Customers call a hosted API and pay per token, or rent a dedicated deployment. The company's own AI agents profile each workload and keep changing the serving setup (engine, GPU kernels, batching, caching, hardware) to make answers come back faster and cheaper. **Why it worked.** Running a large model fast needs performance engineers, and the founders write that only a few hundred people in the world can deeply optimize accelerator hardware. Wafer tries to do that work with AI agents, on AMD as well as NVIDIA GPUs, and to keep doing it after a deployment goes live instead of tuning it once. Evidence grades: A = disclosed by a public company, a filing or a payment processor, or verified by one; B = two or more independent reports agree, or a journalist confirmed it; C = a founder claim that no one else verified; D = reported, not verified: one founder claim or a third-party estimate with no second source; E = no revenue figure found, so the headline is a non-money fact and the other facts carry their own grades; calc = computed on this page from graded facts. ## Summary Wafer is an inference provider: a company that runs AI models on its own GPUs and sells the answers through an API, so the customer does not need its own servers. It serves open-weight models, models whose trained parameters are published, so anyone may run them, such as GLM-5.3, Kimi K3 and DeepSeek V4. What it sells on top is continual inference, its name for agents that watch a customer's traffic and keep changing the serving stack (the inference engine, the GPU kernels, small programs that run on the graphics chip, and the batching and caching rules) to cut latency and cost. Emilio Andere and Steven Arellano, friends from the University of Chicago, founded it in 2025 and went through Y Combinator's Summer 2025 batch. The first products were tools for GPU engineers: an open benchmark suite in July 2025 and an editor extension with a profiler in December 2025. In 2026 it moved to selling inference. Its seed post of April 2026 was signed by five people. The company states a US$4 million seed in April 2026 and a US$40 million Series A on 1 September 2026. The headline is grade D. The Lean AI Leaderboard, a public spreadsheet of small AI companies, lists Wafer with US$8 million in annual revenue and 8 employees for a September 2026 snapshot, and cites a paywalled article in The Information of 1 September 2026 that could not be read here. No payment processor, filing or company statement confirms the figure. By November 2026 the company's team page names 18 people. ## Pricing Wafer serverless prices in November 2026, read from the public model catalog at pass.wafer.ai/v1/models (grade C); dedicated deployments have no published price | Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit | |---|---|---|---| | GLM-5.3-Flash, serverless | | | US$0.05 input and US$0.20 output per million tokens | | DeepSeek-V4-Flash-0731-Fast, serverless | | | US$0.13 input and US$0.23 output per million tokens | | GLM-5.3, serverless | | | US$0.20 input and US$1.80 output per million tokens | | DeepSeek-V4-Pro, serverless | | | US$0.28 input and US$3.96 output per million tokens | | Kimi-K3, serverless | | | US$0.60 input and US$15.00 output per million tokens | | Kimi-K3.1-Pro, serverless | | | US$2.00 input and US$10.00 output per million tokens | | Dedicated deployment | custom | | configured per customer model and traffic; no public price | Serverless use is paid per token from prepaid credits; there is no monthly plan, so the per-month columns are empty. The catalog gives list prices in US cents per million tokens; this table converts them to US dollars. Cached input is cheaper (for example US$0.18 per million tokens for GLM-5.3). The documentation says prices change as the catalog changes, and that accounts may have their own rates or dynamic pricing. Dedicated deployments are quoted per customer. ## How much revenue does Wafer make? The numbers, graded A two-founder company from 2025 that a public leaderboard lists at US$8 million a year in revenue in September 2026, one source only | Fact | Value | Grade | Note | Sources | |---|---|---|---|---| | Annual revenue, September 2026 snapshot | US$8,000,000 | D | Lean AI Leaderboard, column 'Annual Revenue', snapshot 'Sep 2026', with The Information's article of 1 September 2026 as its source. The article is behind a paywall and was not read. One source. | [1, 2] | | Revenue per month implied by that figure | about US$666,667 | calc | Calculated: US$8,000,000 divided by 12. Only as good as the leaderboard figure. | [1] | | Employees at the snapshot | 8 | D | Lean AI Leaderboard, same row. The leaderboard also lists revenue per employee of US$1,000,000 and '4' in its column 'Months to reach ARR'. | [1] | | Team, November 2026 | 2 founders and 16 more people | C | Team page: founders Emilio Andere (CEO) and Steven Arellano (CTO), 10 technical staff and 6 people in sales and operations. Y Combinator's page gives a team size of 15. | [7, 8] | | Team at the seed round, April 2026 | 5 people | C | The seed post is signed 'Emilio, Steven, John, Ian, Danny'. | [6] | | Founded | 2025, Y Combinator Summer 2025 batch | C | Company's information page and Y Combinator's company page. | [4, 8] | | Seed round | US$4 million led by Fifty Years, announced 14 April 2026 | C | Company's seed post and information page. | [6, 4] | | Series A | US$40 million co-led by Marathon and Chemistry, announced 1 September 2026 | C | Company's post. Wing, AMD Ventures and Outset Capital took part; Fifty Years and Y Combinator invested again. The leaderboard's total funding of US$44 million agrees with the two rounds. | [5, 1] | | Valuation at the Series A | US$220 million post-money | C | Company's information page, updated 31 October 2026. The Information's headline says 'US$200 Million-Plus'; the leaderboard lists US$200,000,000. | [4, 2, 1] | | Acquisition offers | reported, no buyer or amount public | D | Only the headline of The Information's article: 'Lands Acquisition Offers'. The text was not read. | [2] | | Tokens served | more than 10 trillion | C | Home page: '10T+ tokens of continual inference'. No date or period is given. | [3] | | Serverless price range | US$0.05 to US$2.00 input and US$0.20 to US$15.00 output per million tokens | C | Public model catalog, read 6 November 2026, 12 models, all marked serverless-only. | [11] | | Hardware | NVIDIA and AMD GPUs | C | Information page. Blog posts report runs on AMD MI355X GPUs. | [4, 9] | | Example customer result | 379 ms average model latency for Y Combinator's voice simulator | C | Wafer's case study, 4,168 model turns; 31% lower than the OpenAI setup and 44% lower than the Cerebras setup that YC tested. Customer measurements reported by Wafer. YC is also an investor. | [12] | | Time from first public release to the revenue figure | 14 months | calc | Calculated: July 2025 (first blog posts) to September 2026. | [9, 1] | How to read this table: a grade D value comes from one source that no one else confirms, here a public leaderboard that cites a paywalled news article. A grade C value is the company's own statement on its pages. A "calc" value was computed on this page from the rows above it. No value is grade A or B. ## When did Wafer launch, and how did it grow? Timeline From two university friends in Y Combinator to a US$40 million Series A in about 14 months. Each entry carries its grade. - **2025** (grade C [4, 6, 8]): Emilio Andere and Steven Arellano, who met as first-year students at the University of Chicago, found Wafer and join Y Combinator's Summer 2025 batch. - **14 July 2025** (grade C [9]): First blog posts, including Chip Benchmark, an open suite that compares open-weight models across accelerators. - **19 December 2025** (grade C [10]): Launch of the Wafer extension for VS Code and Cursor: GPU profiling with NVIDIA Nsight Compute, a compiler explorer and GPU documentation in the editor. - **14 April 2026** (grade C [6]): US$4 million seed round led by Fifty Years. The post is signed by five people. - **10 June 2026** (grade C [9]): Post with DigitalOcean on speeding up Kimi 2.5, DeepSeek V3.2 and GLM-5 on AMD GPUs. - **1 September 2026** (grade C [5, 2]): US$40 million Series A co-led by Marathon and Chemistry. The same day The Information reports a valuation above US$200 million and acquisition offers. - **September 2026** (grade D [1]): The Lean AI Leaderboard lists US$8 million in annual revenue and 8 employees. - **November 2026** (grade C [7, 11]): The team page names 18 people; serverless prices for 12 open-weight models are public. ## Why did Wafer work? ### Agents do the performance work The founders' seed post says the scarce resource is performance engineers who profile, read traces and rewrite kernels by hand. Wafer's agents profile a workload, find the bottleneck and test changes to batching, scheduling, memory layout, caching and decoding, and check each change for correctness before it goes live. ### Keep tuning after launch The company's information page says most optimization is done once before deployment. Wafer repeats it as traffic, models and hardware change, and calls this continual inference. ### Hardware beyond NVIDIA The blog reports fast runs on AMD MI355X GPUs and a joint post with DigitalOcean; AMD Ventures joined the Series A. Getting good speed on cheaper or less used chips is where hand tuning is most scarce. ### Tools first, then inference The company first published a benchmark suite and an editor extension for GPU engineers, then sold the same skill as an inference API with per-token prices and dedicated deployments. ## How it was built These details come from the company's own pages and blog. They are grade C. - Serverless API compatible with the OpenAI and Anthropic request formats, at pass.wafer.ai/v1, paid from prepaid credits. - Models are open-weight models that Wafer hosts on its own fleet, for example GLM-5.3 in NVFP4 format with a 1-million-token context window. - Optimization covers the model configuration, the inference engine, GPU kernels and hardware choice, on NVIDIA and AMD GPUs. - Zero data retention is the default for every serverless request, according to the information page. - Earlier products: an open chip benchmark suite (July 2025) and an editor extension for VS Code and Cursor that runs NVIDIA Nsight Compute profiles (December 2025). Sources: [4, 11, 9, 10] ## Who tried this before Wafer, and why did they fail? Four companies in our database of failed startups sold GPU compute or AI hardware and lost to larger providers. The values below are from that database and are not independently verified. The failed startups come from our database of dead startups, retold here in our own words; Its entries are AI-assisted summaries of public sources, so they carry no evidence grade. ### Vectordash (2019 to 2022; funding US$2M; cause: Competition) - Promise: Cheap GPU power for gamers and AI developers from a network of idle GPUs. - Why it died: Large cloud providers offered integrated GPU services; the spread-out network could not give steady performance. - Record: ### FloydHub (2017 to 2020; funding US$2M; cause: Competition) - Promise: A managed cloud for training machine learning models on GPUs. - Why it died: AWS SageMaker and similar services from the large clouds offered more at scale, and the GPU costs did not leave a margin. - Record: ### Vathys (2018 to 2021; funding US$2M; cause: Competition) - Promise: New processors built for AI workloads, more efficient than GPUs. - Why it died: It could not raise the capital that chip production needs against NVIDIA and other large players. - Record: ### Surge Cloud (2018 to 2023; funding US$1M; cause: Ran Out of Cash) - Promise: A regional cloud for Central and Eastern Europe with lower latency than AWS or Azure. - Why it died: Hardware costs grew with each customer while AWS opened European regions and cut prices. - Record: ## What changed All four sold raw compute or new chips and competed with the large clouds on price and scale. They lost because GPUs and data centers cost the same for them as for anyone, and the large providers bought more cheaply. Wafer sells the tuning, not only the GPU hours. Its claim is that AI agents can do the scarce performance-engineering work, so a small team gets more tokens per GPU than a provider that tunes once by hand. Open-weight models from several labs give it a product to serve without owning a model. Whether this holds against Together AI, Fireworks, Groq or Cerebras, which the company itself compares against, is not shown by any figure on this page. ## Can one person copy Wafer now? **Verdict: Not by one person.** The company raised US$44 million by its own statement and pays for a GPU fleet, and its team is built from kernel and infrastructure engineers. A solo founder cannot copy the inference business. The smaller part, tools for GPU engineers such as profilers inside the editor, is closer to what one person can build. 1. Pick a scarce, measurable skill (here kernel tuning) and build an agent that does part of it. 2. Publish open benchmarks and developer tools first, to show the skill in public. 3. Sell the result as a metered API with public per-token prices, and publish customer latency numbers. 4. Work on hardware that the market leader neglects, where hand tuning is rarest. **What is still open.** Tuning agents for one narrow stack: a single accelerator family, one inference engine, or one workload type such as voice agents, sold to teams that run their own models. ## What could go wrong for Wafer? - **What is not verified.** The US$8 million in annual revenue and the 8 employees rest on one source: the Lean AI Leaderboard row with a September 2026 snapshot, which cites an article by Stephanie Palazzolo in The Information of 1 September 2026. That article is behind a paywall; only its headline, date and summary were read, and they do not show the revenue figure. The company has not published a revenue figure. A reader could look for the full article, a company statement or an investor's disclosure. - **Team size.** Two co-founders started it in 2025, and five people signed the seed post in April 2026. The leaderboard lists 8 employees at the September 2026 figure. In November 2026 the team page names 18 people and Y Combinator's page gives 15, so the company is now larger than the site's small-team range, after US$44 million in funding by the company's statement. - **Funding and acquisition offers.** The rounds and the US$220 million valuation are the company's own statements. The Information's headline says the company 'Lands Acquisition Offers'; no buyer, amount or answer is public. - **Competition.** Wafer's own site has comparison pages against Modal, Cerebras, Together AI, Groq, Baseten and Fireworks, all larger inference providers. On this site Morph also sells fast models through an API, and LLM Gateway and Opper route requests to many model providers without hosting the models themselves. - **Dependence on model providers and hardware.** Wafer does not train its own models. Its catalog lists open-weight models from other labs (GLM from Z.ai, Kimi, DeepSeek, Qwen), so its product depends on those labs continuing to publish weights that customers want. It also depends on GPU supply from NVIDIA and AMD; AMD Ventures is an investor and the blog's speed results are mostly on AMD chips. - **Customer results and testimonials.** Latency and cost figures on the home page and in the Y Combinator case study are customer measurements reported by Wafer. Y Combinator is both a customer and an investor. The YC product in the case study lets users talk to AI versions of YC partners; Wafer only supplies the model answers. The site also publishes dozens of question-style blog pages, such as 'which inference providers are fastest for voice agents', that recommend Wafer. ## Sources 1. Lean AI Leaderboard, public spreadsheet of lean AI companies, row 'Wafer', snapshot September 2026 — https://leanaileaderboard.com/ (grade D). Read the Wafer row in the published Google Sheet that the page embeds, downloaded as CSV on 6 November 2026. A third-party list; it cites The Information for this row. 2. The Information, 1 September 2026, by Stephanie Palazzolo: Wafer, An Inference Provider That Uses Non-Nvidia Chips, Lands Acquisition Offers and $200 Million-Plus Valuation — https://www.theinformation.com/newsletters/ai-agenda/wafer-inference-provider-uses-non-nvidia-chips-lands-acquisition-offers-200-million-plus-valuation (grade D). Could not read beyond the paywall. Read the headline, date, author and summary in the page's metadata only. 3. Wafer home page: Continual Inference — https://www.wafer.ai/ (grade C). Read in full on 6 November 2026. Source for the token count and the customer quotations. 4. Wafer: Official information about Wafer, last updated 31 October 2026 — https://www.wafer.ai/ai-info (grade C). Read in full. Source for the founding year, founders, rounds, valuation, products and hardware. 5. Wafer blog, 1 September 2026: Wafer Raises $40M Series A to Build AI That Optimizes AI — https://www.wafer.ai/blog/series-a (grade C). Read in full. 6. Wafer blog, 14 April 2026: Announcing our Seed Round — https://www.wafer.ai/blog/seed-round (grade C). Read in full. Source for the founders' story, the seed amount and the five signatures. 7. Wafer team page — https://www.wafer.ai/team (grade C). Read in full on 6 November 2026. Counted 18 named people. 8. Y Combinator company page: Wafer, AI that makes AI fast — https://www.ycombinator.com/companies/wafer (grade C). Read the page data: batch Summer 2025, founded 2025, team size 15, two founders. 9. Wafer blog index — https://www.wafer.ai/blog (grade C). Read the list of post titles and dates, July 2025 to July 2026. 10. Wafer blog, 19 December 2025: Introducing the Wafer Extension for VS Code and Cursor — https://www.wafer.ai/blog/wafer-extension (grade C). Read the opening sections. 11. Wafer model catalog (pass.wafer.ai/v1/models) and documentation page Models and Capabilities — https://docs.wafer.ai/serverless/models.md (grade C). Read the documentation in full and the public catalog JSON on 6 November 2026, without an API key, so list prices. 12. Wafer case study: How YC lowered response latency and increased time spent with its AI partners — https://www.wafer.ai/cases/yc (grade C). Read in full. Customer measurements reported by Wafer. 13. Database of failed startups, local export — /method/#dead (grade C). Local export. AI-written summaries, not independently verified. Source for the failed predecessors. Contact: aiolum@livyden.com. Run Wafer? Send corrections or a verified figure to aiolum@livyden.com (subject: Correction: Wafer).