doubleAI, the Tel Aviv startup co-founded by Mobileye CEO Amnon Shashua, has released its debut product, a "code execution model" called Maxwell that is meant to make existing software run faster on GPUs and CPUs. According to the Wall Street Journal's October 19, 2026 report as relayed by Quartz and the Swedish outlet Borsvarlden, the company has raised $220 million and plans to charge customers a share of the compute savings it delivers. The headline number, a 67% cost cut for an early customer, is doubleAI's own claim.
That makes this a classic claimed-versus-verified story. The product category is real and important, since compute is the main cost of AI, but the numbers come from the vendor. Below we separate what is reported, what is only claimed, and how readers can judge it. For the hardware backdrop, see our coverage of B300 rental prices at record highs and AWS raising GPU Capacity Block prices.
Quick facts
| Question | Answer | Source |
|---|---|---|
| What is Maxwell? | doubleAI's first "code execution model" for optimizing GPU and CPU performance | WSJ via Quartz, Borsvarlden |
| Who runs doubleAI? | CEO Amnon Shashua (also Mobileye CEO) and CTO Shai Shalev-Shwartz | Company background reporting |
| Funding | $220M total; investors include Lightspeed, Bessemer, Nvidia, Dell | WSJ via Quartz |
| Hardware | About half the raise reportedly spent on 1,000 Nvidia B300 chips | WSJ via Techmeme summary |
| Pricing | A share of customer savings | WSJ via Borsvarlden |
| Headline claim | 67% lower compute cost for an early customer | doubleAI, unverified |
| Customers named? | No | Reports |
| Independent benchmark? | None published; a second-hand report cites LTX seeing 2x faster inference | Our search |
What is a code execution model?
Most AI models read or write text. doubleAI describes a code execution model as something that sits inside the running program, works autonomously and finds ways to improve how the code performs, according to the report. In plain terms: instead of a person hand-tuning GPU kernels, an AI system rewrites the performance-critical code and checks that the result still behaves the same.
That second part, equivalence, is where the difficulty lies. Shashua told an Israeli conference, per Calcalist, that "debugging something that 'cheats' like a machine is extremely difficult" and called verification a major open challenge in AI-assisted work. doubleAI's pitch rests on being able to verify rewrites, and we found no public description of its method in the sources we could read, so that is the hardest part to confirm from outside.
GPU chiplets on one substrate, showing the hardware that code execution models target for optimization
The earlier proof point: WarpSpeed and cuGraph
Maxwell did not come out of nowhere. In March 2026 doubleAI published WarpSpeed, a system that rewrote CUDA kernels in Nvidia's cuGraph library, part of the RAPIDS ecosystem. As reported by TechTime, the company claimed a 3.6x average speedup over versions refined by Nvidia's engineers over about a decade, with every tested kernel improved and more than half gaining over 2x. The optimized code was released on GitHub so users could deploy it without changing application code. A later article describes doubleGraph as a free, verified replacement for cuGraph, according to RuntimeWire.
That TechTime article is candid about the limits: it does not independently verify the claim, describes no benchmark methodology, and says the gains would need to "hold up under community scrutiny." It is a sensible frame for Maxwell too. The cuGraph work is the best public evidence that the approach works, because the code is open and anyone with a GPU can rerun it, which a private customer case cannot offer.
What doubleAI says about its approach
Shashua has argued that recent AI wins in competitions rest on three "hidden crutches": abundant training data, easy verification and short reasoning chains. GPU optimization has none of them, he says, and he claims even advanced coding agents "fail dramatically" at it. The company brands its aim as Artificial Expert Intelligence, narrow high-value expertise instead of general intelligence. Compare that thesis with other attempts to make agents cheaper or leaner, such as Nvidia's EdgeWise delegation scheme that cut agent cost 62% and Strata running a 125B-class model on a consumer GPU. Those reduce cost by changing the model; doubleAI says it reduces cost by changing the code the model runs on.
Claimed versus verified
A figure beside a ruler with a green mark, representing measuring a vendor's compute savings claim against a baseline
| Claim | Who says it | Status |
|---|---|---|
| Maxwell exists and is released | WSJ via Quartz, Borsvarlden | Reported by multiple outlets; we could not open the WSJ original |
| $220M raised | WSJ via outlets | Consistent across the reports we found; no filing seen |
| 1,000 B300 chips bought | WSJ via Techmeme summary | One-line summary only; no detail checked |
| 67% compute savings for an early customer | doubleAI | Unverified; customer unnamed, baseline undisclosed |
| 3.6x average speedup on cuGraph | doubleAI | Company-published; code on GitHub for replication; no independent study found |
| Verification of rewrites | doubleAI | Method not public in the sources we read |
| Charge a share of savings | doubleAI | Stated plan, not a published price sheet |
| LTX reportedly saw inference for its own model run twice as fast | Search-result reports of the WSJ story | Second-hand; we did not see LTX's own statement |
Why pricing as a share of savings matters
Charging a percentage of savings sounds fair, but it moves the argument to the baseline. If a customer's code was already poorly tuned, a 67% reduction is easier than if it was running well. A buyer should agree in advance on how the "before" cost is measured: same hardware, same inputs, same latency targets, and a correctness test suite. Also decide who pays if an optimized kernel gives a subtly different numerical result, which matters in training and in scientific code.
A coin stack shrinking step by step with a green arrow down, illustrating compute cost reduction
How to evaluate a tool like this yourself
- Ask for the baseline. Which library version, compiler flags, GPU and batch size produced the "before" number?
- Run your own test suite against the rewritten code, including edge cases and numerical tolerances.
- Profile before and after. Compare kernel time, memory bandwidth and end-to-end latency, not just one microbenchmark.
- Check portability. A rewrite tuned for a B300 may not help on older cards; ask how it behaves across your fleet.
- Read the licence and data terms for code that leaves your environment.
- Pilot on one workload and cap the savings share until results repeat.
Why it matters
If rewriting hot loops by machine works as described, it attacks cost from a different direction than cheaper chips or smaller models. It also touches the supply story: a startup that spends half its raise on 1,000 B300 chips is betting that optimization is compute-hungry too. Our posts on the Taiwan B300 smuggling indictments and Nvidia's plan to turn GPUs into an asset class show how tight that supply is, and how much value would come from squeezing more work out of each chip.
At the same time, Nvidia is both an investor in doubleAI and the owner of the libraries doubleAI claims to have beaten. That is not a conflict by itself, but readers should note it when weighing the story.
What this means for what you build or pay
For a team renting GPUs, the practical question is whether a vendor can lower the bill without a model swap. Today the answer is "maybe, and prove it." If you run stable, hot code paths such as graph analytics, data preprocessing or custom inference kernels, a rewrite that holds numerical results could pay for itself quickly under a savings-share contract. If your workload is mostly calls to a hosted model API, none of this applies, because you do not control the kernels.
For individual developers, the lesson is narrower. Coding agents are good at writing application logic, but low-level performance work needs fast feedback loops, hardware access and strict correctness checks. That is why doubleAI says it bought its own B300 fleet: tuning needs real silicon to test on. Teams already experimenting with agents that profile and rewrite code should keep tests and benchmarks automated, so any vendor claim can be checked against your own numbers within hours.
Common questions people are asking
Is this the same as a compiler? Not exactly. Compilers apply fixed transformations; the company describes a model that proposes new rewrites and then verifies them. Whether it beats an expert-tuned compiler flag set on your code is the open question.
Does it replace performance engineers? The reports do not say so. The sensible reading is that it automates the repetitive search part and leaves humans to define goals, baselines and acceptable risk.
Is Nvidia threatened? Faster software can mean fewer chips per job, but Nvidia is also an investor, and the reports give no sign of a dispute.
What to watch next
- A named customer, or a case study with a disclosed baseline.
- Open code or a technical report for Maxwell comparable to the WarpSpeed release.
- Independent reruns of the cuGraph kernels.
- Pricing details: the percentage, the cap and the measurement rules.
- Any Nvidia response on the cuGraph comparison.
Related reading
- B300 rental prices at a record: what the indexes show
- AWS GPU Capacity Blocks price hike
- Taiwan indicts 9 over smuggled B300 servers
- Nvidia's $500B GPU asset-class plan
- Nvidia EdgeWise cuts agent cost 62%
- Strata: a 125B-class model on a consumer GPU
Figures are accurate as of October 19, 2026 and come from press reports of company statements; we could not open the original Wall Street Journal article.