Sell or Build: Wafer's Refusal to Sell
A practical rule for AI infrastructure founders: prove a benchmark, own the workload, then raise from the buyers who wanted to acquire you.
An acquisition offer can land before the founder has a public number. Wafer rejected purchase proposals from several cloud and inference firms.
The benchmark is the offer
In AI inference, the buyer often sees a gap before the founder can state it: your stack moves tokens faster, cheaper, or more reliably than the incumbent. That gap is the asset. If you cannot put it in one sentence, the offer will set the price.
Wafer said its agents enabled AMD's MI355X to deliver roughly 80 percent of the inference throughput of Nvidia's B200 while costing less than half as much for select models such as GLM-5.2.
A number like that changes the negotiation. It turns GPU optimization from a pitch into a measured workload, and it forces the buyer to explain why it should own the workload instead of buying the company that produced it.
For open-source models, the serving stack often decides the margin. The model weights are fixed, so the founder's edge comes from scheduling, batching, memory layout, and hardware choice. A clean benchmark makes that edge visible to customers, not just to acquirers, and gives the founder a repeatable way to answer the next technical question.
Own the workload before you own the exit
The strongest objection is that independence is a luxury. A clean exit removes execution risk, hiring risk, and the chance that a larger lab changes the model mix. That fear is real. But in AI infrastructure, the buyer's offer often prices the current gap, not the workload you can own. The exit can be the right answer; the mistake is treating it as the default before the workload is measured.
Wafer generated $8 million in annual recurring revenue roughly three to four months after its inference cloud launch, while its team was about seven people during peak scaling.
That ARR milestone shows the workload is sticky enough to pay for itself and small enough for a tight team to run. It gives the founder a number that does not depend on a buyer's internal model. It also makes the workload easier to price, because the team can see which models and customers drive the margin.
Revenue also changes leverage. A buyer can still offer a multiple of current ARR, but the founder can compare that multiple against the cost of continuing to serve the workload. If the team can keep the margin, the exit is one option, not the only path.
Raise the buyers, not just the round
Once the benchmark and revenue are public, the funding round becomes a test of who wants the workload. Wafer completed a $40 million Series A on September 1, 2026, at a valuation above $200 million, roughly 50 times its April 2026 $4 million seed, which was led by Fifty Years with Jeff Dean and Wojciech Zaremba participating as early angels.
Marathon Management Partners and Chemistry co-led the Series A, with AMD Ventures, Wing Venture Capital, and Outset Capital also participating.
The investor mix matters because the parties closest to the workload saw the same gap the founders measured. It also gives the founder a clearer basis for the next raise, and a reason to expect the next offer to be tested against measured usage.
The order is the leverage
- Prove the benchmark before you discuss price. Put the cost or throughput number in front of a customer, a public post, or a technical report. If the number cannot survive scrutiny, the offer will define the ceiling.
- Own the workload before you own the exit. Build the inference cloud, the API, the billing, and the support loop. Make the workload yours, not just your demo.
- Raise the buyers after the workload is real. Strong capital often comes from the parties who would have acquired you, because they already understand the gap and the risk.