doubleAI Raises $200M for AI That Rewrites Code as It Runs
doubleAI Raises $200M for AI That Rewrites Code as It Runs
Update, Oct. 20: WSJ puts total raised at $220M
The Wall Street Journal reports doubleAI has actually raised $220M in total (vs the $200M Series A headline) — and half the round will buy 1,000 Nvidia B300 Blackwell GPUs (Daily AI Brief, citing WSJ). The pricing backdrop: SemiAnalysis notes one-year H100 contract rentals are up 60% in a year — the iron price of the AI buildout that doubleAI's "code execution models" are trying to shrink.
Every AI company on the planet is begging for more chips. Amnon Shashua just raised $200 million on the opposite bet: that the chips we already own are being colossally wasted.
The Tel Aviv startup, doubleAI, announced Monday that it has closed a $200 million Series A co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with Nvidia, Dell Technologies Capital, Greenfield, Pitango and BRM participating. The round values the roughly 40-person company at more than $1 billion, according to CTech, and brings its total funding to about $220 million.
What it sells is stranger than the valuation. doubleAI calls its technology "Code Execution Models," or CEMs — AI that lives inside running software and rewrites the machine instructions underneath it, while the program keeps running. Its first model, Maxwell, is now open for early access. And the company plans to charge customers not by the seat or the token, but by the dollar of compute it saves them.
Server racks at NOIRLab's headquarters in Tucson, Arizona — the kind of compute hardware doubleAI's Maxwell aims to make run more efficiently. Photo: NOIRLab/NSF/AURA/T. Slovinský, via Wikimedia Commons (CC BY 4.0).
The man who sold Mobileye to Intel is done writing code
Shashua does not need the money, and he does not need the introduction. He co-founded Mobileye in 1999 with Ziv Aviram, took it public, and sold it to Intel for $15.3 billion in 2017 — still the largest exit in Israeli tech history. He went on to co-found OrCam and AI21 Labs, and collected the Israel Prize along the way.
In July, Mobileye said Shashua would step down as its chief executive once it names a successor. He has also left his roles at AI21, according to CTech. The pattern is familiar: Shashua builds the thing, hands over the keys, and goes looking for a harder problem. This time the problem is the economics of compute itself.
"AI has changed how software is written, but not how it runs," Shashua said in the company's announcement. "Code Execution Models change that."
It is a neat line, and it happens to be aimed squarely at the industry's most uncomfortable question: who is actually going to pay for all of this AI? While the rest of the market argues about the industry's $113 billion debt machine and OpenAI's scramble to lock in $30 billion, Shashua is selling the one thing every buyer wants — a smaller bill.
Maxwell works for the machine, not the programmer
Here is the distinction that matters. Coding assistants — Copilot, Cursor, Claude Code — work for the programmer. They write source code, a human reads it, and the machine runs it exactly as written. Maxwell skips the human entirely. It watches a program execute, studies the hardware it runs on and the data flowing through it, then rewrites the compiled binary underneath to fit that exact moment. The source code never changes. The program just gets faster while it runs.
The company describes CEMs as sitting "between the code and the silicon." Each improvement is formally proven correct — "Maxwell formally and mathematically proves that every optimization it unlocks leaves the program's behavior unchanged," the company says — and every proven version becomes the starting point for the next round of optimization. The code, in effect, keeps improving itself for as long as it runs.
The rear of a server rack at NERSC. doubleAI says Maxwell optimizes code while it runs on CPUs and GPUs alike. Photo via Wikimedia Commons (CC0 1.0).
"In other words, they are the first AI that works for the machine, not the programmer," Shashua said. It is the kind of sentence that sounds like marketing until you notice who is paying $200 million to find out whether it is true. Nvidia does not co-invest in many startups whose pitch is "you need fewer of our chips."
The numbers are colossal — and entirely self-reported
doubleAI's evidence file is genuinely startling. Maxwell rewrote Nvidia's own open-source cuGraph graph-analytics library — revealed back in March, according to SiliconANGLE — and sped up 55% of its algorithms by two times or more. It produced an optimized version of pandas, the workhorse Python data toolkit, improving its PDS-H query benchmark performance by a factor of 38. It doubled the video-generation speed of LTX-2.5, Lightricks' open world model. The company further claims data pipelines running 38 times faster and AI inference kernels 4.5 times faster.
We opened doubleAI's research write-up on the LTX-2.5 work and checked the fine print, and the detail is more convincing than the headline numbers. A five-second 4K clip went from 250 seconds to 134 on two B200 GPUs. The output quality measured 41 dB PSNR — past the threshold the company itself cites for visually indistinguishable. Maxwell found more than a dozen changes, from rebalancing how two GPUs exchange attention data to catching a rare kernel hang, and every change was published open source.
Now the uncomfortable part, and it needs saying plainly: every one of these numbers comes from the company selling the product. The press-release benchmark circuit is a shoddy substitute for independent testing, and 38x on a query benchmark is the oldest trick in the vendor playbook — pick the workload where you shine. What keeps this story out of the pure-hype bin is the third party on the record. Zeev Farbman, co-founder and CEO of LTX, confirmed Maxwell doubled his model's inference speed: "Increased speed translates directly to cost savings," he said. "We measure GPU usage time, and since GPU time has a set market price at any given moment, cutting GPU usage time in half directly impacts service delivery costs and margins." One confirmed customer beats ten company benchmarks. The push for leaner inference is the same inference-efficiency race the rest of the industry is running.
Charge me what you save me
The pricing model is the most interesting thing doubleAI announced, and it got the least attention. The company plans to charge customers based on the computing costs it actually eliminates, CTech reports. The Journal's telling of it: cut a client's $10 million compute bill to $6 million, and take a percentage of the $4 million saved. Shashua did not specify the percentage, which is exactly the number that will decide whether this is a business or a stunt.
But the logic is hard to argue with. Software pricing has always been decoupled from value delivered — you pay per seat whether the tool helps or not. Tying the price to verified savings aligns the vendor with the buyer for the first time in enterprise software. Shashua put it in terms the big labs will understand: "There are industries or companies that have more critical challenges than saving $500,000 on a $2 million computing bill, but for major AI labs, a 10% reduction in computing costs could amount to hundreds of millions of dollars."
He expects Maxwell to be fully deployed later this month, and named the likeliest mega-customers without blinking: OpenAI and Anthropic.
Half the money goes straight back to Nvidia
There is a delicious irony buried in the funding details. The Wall Street Journal reported that doubleAI will spend about half of the $200 million round buying 1,000 of Nvidia's B300 Blackwell GPUs. A company whose pitch is "you need less compute" is spending nine figures on compute — because proving you can shrink everyone else's GPU bill requires a mountain of GPUs of your own.
The timing is no accident. SemiAnalysis data cited by the Journal shows the hourly rental cost of Nvidia H100 compute on one-year contracts has risen 60% in the past year. Compute is the scarcest resource in tech, and scarcity is what makes an efficiency play valuable. "More hardware is being built, but the need for compute is rising faster than the ability to build GPUs and CPUs," Shashua told the Journal. "If you can utilize the hardware better, then you can achieve something incredible in terms of savings."
This is the same squeeze Hock Tan warned about when he said AI's biggest problem isn't chips — the bottleneck is what it costs to run them. doubleAI is the first well-funded startup to make that sentence its entire business plan.
The chess record was the audition
Before the funding, there were the stunts — and they were better than stunts. Earlier this month, doubleAI's internal AI research system, which it calls Double Agent, broke a chess record set in 1968: it found a position in which all 63 legal moves give check, topping Norbert Ringeltaube's record of 60 that had stood for 58 years. It also proved no position can do better, and that the position is reachable in a legal game — every step verified with the Lean proof assistant and SAT solvers. The company's LinkedIn announcement closed with a deadpan line: "Chess is the fun part."
Around the same time, an internal agent produced a genuinely new graph-theory result, paired with a Lean verification file. The point of both exercises is the same as the product pitch: this is AI that doesn't just generate plausible output, but proves its work. In a field drowning in demos that fall over in production, "formally verified" is the most unfashionable — and most interesting — claim in the announcement.
Why this one lands in America first
Strip away the Tel Aviv dateline and this is an American story. The money is American — Lightspeed, Bessemer, Dell. The chips are Nvidia's. The customers Shashua is courting are OpenAI and Anthropic. And the pain is most acute in the US, where AI labs are signing compute contracts at prices that would have been unthinkable two years ago.
"Much of the world's compute goes to running code far more slowly than it could," Shashua said. "Every program Maxwell speeds up gives some of that back."
The honest read: doubleAI is a $1 billion bet that the AI boom's scarcest resource isn't intelligence, it's efficiency — and that the man who taught cars to see can teach code to run. The benchmarks are self-reported, the pricing percentage is undisclosed, and Maxwell has yet to prove itself on anyone's production workload except LTX's. But Nvidia just put its own money behind a startup whose mission is to make you buy fewer Nvidia chips. When the chipmaker bets against its own scarcity, pay attention. doubleAI is another compute-efficiency insurgent betting that the next AI edge is efficiency, not scale.
doubleAI FAQ: what Maxwell, the $200M raise and CEMs mean
What is doubleAI?
Tel Aviv frontier lab founded by Mobileye co-founder Amnon Shashua; its "code execution models" (CEMs) optimise software as it runs between code and silicon; raised $200M Series A at $1B+ (WSJ now puts total raised at $220M).
How does Maxwell work?
Maxwell watches running programs, rewrites machine binaries with mathematically proven correctness; CPUs and GPUs; self-improving (each verified version becomes the next baseline).
Who invested in doubleAI?
Lightspeed + Bessemer co-led; Nvidia, Dell Technologies Capital, Greenfield, Pitango, BRM participated.
Sources: doubleAI company announcement via BusinessWire (Oct 19, 2026); doubleAI research post "Maxwell Makes LTX-2.5 Generate 4K Video Nearly 2x Faster"; CTech/Calcalist; SiliconANGLE; The Wall Street Journal (via syndicated reporting); Daily AI Brief; doubleAI LinkedIn.
Daniel Rao covers AI infrastructure, chips, and the economics of compute for TechPulse USA — who pays for the AI boom, and what they get for it.
Comments
Post a Comment