An Israeli startup called doubleAI unveiled a new kind of software on October 19, 2026, that claims to speed up running programs without anyone touching a single line of source code. The product, named Maxwell, watches an application while it executes and rewrites the machine code underneath it in real time. doubleAI says the approach, which it calls a Code Execution Model, pushed real-world data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster, according to the company’s own benchmarks. The announcement, first reported by The Register, lands at a moment when compute costs are becoming one of the biggest line items in enterprise technology budgets.
Maxwell is now open for early access to software developers, and the pitch is unusual enough that it has drawn attention well beyond the usual developer-tools crowd. Instead of asking engineers to rewrite slow code, doubleAI wants its system to sit inside the runtime and fix the performance problem on its own, on whatever CPU or GPU the program happens to be running on that day. This is a runtime code optimization story, but it is also a story about who gets to decide how fast software runs, and whether an AI system can be trusted to change machine instructions without breaking the program it is optimizing.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Maxwell Actually Does
Most tools that claim to speed up code fall into one of two buckets. Ahead-of-time compilers optimize source code before a program ever runs. Just-in-time compilers, the kind found inside the Java Virtual Machine or a browser’s JavaScript engine, generate machine code while a program executes but are built into the language runtime itself and tuned for general cases. Maxwell is pitched as something else: a standalone system that attaches to an already-compiled binary, observes it in production, and rewrites the underlying machine code to match the specific hardware and the specific data flowing through it at that moment.
Technology investor Adam Fisher described the product bluntly in a public post: “Maxwell is the first Code Execution Model.” He went on to explain the mechanism in more detail, writing that “it lives inside a running program on a CPU or GPU, watches the hardware & data, optimizing the machine instructions underneath.” That framing, posted on X, captures the core pitch: Maxwell does not ask a developer to change anything about how an application is written. It changes what happens after the code leaves the compiler.
doubleAI itself has described Maxwell in similarly ambitious terms. In a company research publication, the firm called Maxwell “doubleAI’s artificial expert intelligence for performance engineering,” a description that positions the product less as a point tool and more as a standing layer that sits between an application and the chips it runs on. The company says Maxwell works across both CPUs and GPUs, which matters because the two chip types have very different performance bottlenecks and very different kinds of code that benefit from tuning.
A Formal Proof Behind Every Rewrite
The part of Maxwell’s pitch that tech workers are most likely to push back on is trust. Handing an AI system the ability to rewrite machine code in a live, revenue-generating pipeline is the kind of idea that makes a site-reliability engineer nervous. doubleAI’s answer to that concern is a formal correctness guarantee: the company says every optimization Maxwell applies is formally proven to be mathematically equivalent to the original code before it ships. In theory, that means the rewritten binary should produce identical outputs to the original program, just faster, because the proof step is meant to reject any rewrite that changes behavior rather than simply speed.
That claim has not been independently verified by outside researchers, and doubleAI has not published the formal methods or proof system it uses in enough technical detail for third parties to replicate the verification step. For now, the correctness guarantee is a company statement rather than a peer-reviewed result, which is a distinction worth keeping in mind for any engineering team evaluating the early-access program.
The Performance Numbers doubleAI Is Publishing
doubleAI’s headline figures are a more than 38 times speedup on data pipelines and a more than 4.5 times speedup on AI inference kernels, both measured on what the company describes as real-world workloads. Those are the numbers the company has promoted most heavily since the October 19 launch, and they are the figures cited by The Next Web in its coverage of the announcement.
There is a wrinkle in the company’s own materials, though. Maxwell’s early-access page currently advertises a 4.25 times improvement for AI inference kernels, a figure that does not quite match the 4.5 times number used elsewhere in doubleAI’s promotional materials. The gap is small, and it may simply reflect different test workloads or different points in time as the product has been tuned, but it is a reminder that none of these numbers come from an independent benchmark suite. They are company-reported results from company-selected workloads, and readers should treat them the way they would treat any vendor benchmark: directionally interesting, not independently audited.
doubleAI has also pointed to a specific customer result. One early customer reportedly cut compute costs by 67% after adopting Maxwell, a figure highlighted in coverage from RuntimeWire. The company has not named that customer publicly, so the figure cannot yet be checked against a named deployment, but a 67% reduction in compute spending would be a meaningful number for any finance team watching a cloud bill climb alongside AI adoption.
The LTX Video-Generation Case Study
The most concrete, specific example doubleAI has put forward involves LTX, which the company says doubled its model-inference speed after adopting Maxwell. According to the research write-up doubleAI published, Maxwell cut the time needed to generate a 4K video clip from the LTX-2.5 model on two Nvidia B200 GPUs from roughly 250 seconds down to about 134 seconds, which works out to close to 1.87 times faster rather than a clean doubling. doubleAI attributes part of that gain to moving work that was previously running on the CPU over to the GPU, batching together small kernel launches that were wasting cycles, cutting down on GPU-to-GPU data transfer, and picking better-suited kernels for the specific operations in the pipeline.
That level of detail is useful because it shows the mechanism is not a single magic trick. It is closer to a bundle of classic performance-engineering techniques, the kind that teams working on AI coding agents have had to learn to apply by hand, now applied automatically and continuously by a system that can watch the workload change over time.
| Metric | Reported Result | Source |
|---|---|---|
| Data pipeline speedup | More than 38x faster | doubleAI company benchmarks |
| AI inference kernel speedup (headline figure) | More than 4.5x faster | doubleAI company benchmarks |
| AI inference kernel speedup (early-access page) | 4.25x faster | heymaxwell.ai early-access page |
| LTX-2.5 4K video generation (2x Nvidia B200) | 250s to 134s per clip (about 1.87x) | doubleAI research publication |
| Compute cost reduction, one early customer | 67% cut | Reported by RuntimeWire |
| Hardware coverage | CPUs and GPUs | doubleAI |
| Availability | Early access for developers | doubleAI |
Who Built Maxwell
doubleAI’s two named co-founders are Amnon Shashua and Shai Shalev-Shwartz. Shashua is widely known in the autonomous-vehicle world as the co-founder and chief executive of Mobileye, the Jerusalem-based computer-vision company behind driver-assistance chips used across much of the global auto industry. Shalev-Shwartz has a long academic and industry background in machine learning. Bringing that pairing of autonomous-systems and machine-learning expertise to a performance-engineering startup is an unusual move, but it fits a broader pattern of veteran AI researchers pivoting toward infrastructure problems once the model-building side of AI started to feel like the easier half of the industry’s challenge.
Shashua has framed the motivation behind Maxwell in economic terms. In comments reported from an interview, he said: “More hardware is being built, but the need for compute is rising faster than the ability to build GPUs and CPUs.” That line gets at the core bet behind doubleAI’s business: if chip supply cannot keep pace with AI demand, then squeezing more useful work out of every chip that is already deployed becomes just as valuable as buying more chips. That is the same pressure driving a wave of coverage this year on component shortages tied to AI demand and on data-center operators scrambling for power and silicon.
Inside the Early Access Program
doubleAI has opened Maxwell to software developers through an early-access program rather than a full public launch, a rollout pattern common among infrastructure startups that need real production traffic to validate a product before scaling it broadly. Early access programs for runtime tools like this typically start with a narrow set of workloads, usually data pipelines and inference serving, since those are the categories where doubleAI has already published results. The company has not detailed public pricing, enrollment limits, or a timeline for general availability, so for now the early-access page at heymaxwell.ai is the only official channel for developers who want hands-on access to Maxwell.
What is notable is how directly doubleAI is targeting infrastructure and platform teams rather than individual application developers. A tool that rewrites binaries at runtime is not something most teams would bolt onto a hobby project. It is pitched at organizations already running large-scale data pipelines or AI inference workloads on a fleet of CPUs and GPUs, where even a modest percentage improvement in throughput translates into a meaningful dollar figure on a monthly cloud invoice.
How Maxwell Compares to Existing Optimization Tools
To understand what is genuinely new about Maxwell, it helps to place it next to the optimization tools engineering teams already use. Gal Beniamini, a co-founder at doubleAI, offered a direct comparison to the technology most people already understand: just-in-time compilers. In an email statement to The Register, Beniamini said: “Both Maxwell and a JIT compiler work while the program runs and both produce new machine code.” That is the similarity. The difference, according to doubleAI, is scope. A JIT compiler is built into a specific language runtime, optimizes within the rules of that runtime, and generally does not carry a formal proof of equivalence for every optimization it applies. Maxwell is pitched as runtime-agnostic, hardware-aware across both CPU and GPU targets, and backed by the formal verification claim described above.
Ahead-of-time compilers sit at the other end of the spectrum. Tools built around compiler infrastructure such as LLVM can apply aggressive optimizations, but they do so once, before the program ever sees real traffic or real data, which means they cannot react when the actual workload in production looks different from what the compiler assumed at build time. Manual performance engineering, the kind teams have long relied on with profilers and hand-tuned kernels, can be extremely effective but is slow, expensive, and has to be redone every time the hardware or workload shifts. Maxwell’s pitch is that it collapses the gap between those two approaches: continuous, data-aware tuning without a human in the loop for every change.
| Approach | When It Optimizes | Changes Source Code? | Formal Correctness Proof? | Typical Use Case |
|---|---|---|---|---|
| Ahead-of-time compiler (e.g. LLVM-based toolchains) | Build/compile time | No, works from source | No | General application builds |
| JIT compiler (e.g. JVM HotSpot, V8) | Runtime, within a managed language | No | No | Managed-language execution |
| Manual profiling and hand optimization | Developer-driven, periodic | Yes | No | Targeted hot-path fixes |
| doubleAI Maxwell (Code Execution Model) | Runtime, continuous | No, binary only | Yes, according to doubleAI | Data pipelines, AI inference kernels |
Historical Context: The Long Road to Automated Performance Tuning
The idea of software that optimizes itself while running is not new, but most prior attempts have been narrow. Dynamic recompilation dates back decades in the world of managed runtimes, and profile-guided optimization has let compilers use data from previous runs to make better decisions for future builds since the 1990s. What has been missing until now is a general-purpose system that treats runtime optimization as its own product category, separate from any specific language or framework, and backs its changes with a correctness proof rather than relying on a human reviewer to catch regressions.
That gap has become more expensive to leave open as AI workloads have scaled. Inference kernels running large language models and video-generation systems chew through GPU time at a cost that was almost irrelevant five years ago and is now a board-level budget line. The same pressure has shown up elsewhere in this year’s coverage of the AI hardware squeeze, including reporting on how AI-assisted code review is reshaping established software projects and on chipmakers raising prices as demand outpaces supply. Maxwell arrives into that environment less as a novelty and more as a direct response to a cost problem that has been building for a while.
Why This Matters for the AI Coding Tools Market
Most of the attention on AI coding tools over the past two years has focused on code generation: assistants that write new functions, suggest fixes, or review pull requests. Maxwell is pointed at a different part of the software lifecycle entirely, the part that happens after code has already shipped and is running in production. If a Code Execution Model can reliably deliver double-digit speedups on existing workloads without a single line of source code changing, it opens up a new lane in the coding-tools market that sits alongside, rather than competing directly with, tools like AI coding agents built for writing and reviewing code and IDE-integrated assistants from vendors such as JetBrains.
That distinction matters commercially. A coding assistant has to earn trust from an individual developer one pull request at a time. A runtime optimization layer has to earn trust from an infrastructure team all at once, because it sits in the production path of every request the optimized service handles. That is a much higher bar to clear, and it is also why doubleAI is leading with a formal verification claim rather than simply publishing a benchmark chart. The product only has a market if engineering leaders believe the correctness guarantee holds under real load, not just in a demo.
It is also worth noting how this plays against the broader trend of AI agents taking on more of the software development lifecycle. Tools like Cursor’s AI agents, which now merge a majority of pull requests on some teams, have shown that developers are increasingly comfortable letting AI systems make direct changes to a codebase, provided there is a review step in the loop. Maxwell removes even that review step from the optimization side of the equation, betting that a formal proof can substitute for human review when the change in question is purely about performance rather than functionality.
The Compute Cost Problem Behind the Launch
Maxwell did not appear in a vacuum. Enterprise cloud bills tied to AI workloads have climbed sharply over the past two years as companies moved from experimenting with large models to running them at scale for real users. GPU capacity has been the tightest constraint in that equation, with reports throughout 2026 pointing to extended wait times for high-end accelerators and rising prices for the memory and storage that support them. doubleAI’s framing, through Shashua’s comments about compute demand outpacing hardware supply, taps directly into that pressure. If a company cannot get more GPUs fast enough, getting more useful work out of the GPUs it already has becomes the next best lever, and that is precisely the gap doubleAI is positioning Maxwell to fill.
That framing also explains why doubleAI is leading its public messaging with a cost-reduction figure, the reported 67% cut in compute spending for one early customer, rather than leading purely with a speed figure. A chief financial officer evaluating cloud spend cares less about an abstract multiple on throughput and more about what shows up on the next invoice. Positioning Maxwell as a cost tool as much as a performance tool widens its potential buyer base beyond engineering teams to include finance and operations stakeholders who sign off on infrastructure budgets.
Open Questions and Risks
Several questions remain unanswered after the October 19 launch. doubleAI has not published independent, third-party benchmark verification of its 38x and 4.5x figures, so the numbers currently rest on the company’s own testing methodology. The firm also has not disclosed the identity of the customer behind the reported 67% compute-cost reduction, which makes that figure hard to contextualize against a specific workload or industry.
- How does Maxwell behave when a workload changes shape dramatically between runs, rather than drifting gradually?
- What happens to the formal proof chain when the underlying hardware itself receives a firmware or driver update?
- Will doubleAI publish enough technical detail about its verification methodology for independent security researchers to audit it?
- How does pricing work once early access ends, and will the model scale economically for smaller engineering teams rather than only large AI infrastructure operators?
- Does rewriting binaries at runtime introduce any new attack surface that security teams need to monitor, separate from the correctness question?
That last point deserves attention from any team that works on the security side of AI-assisted software, since a system with the ability to alter machine code in a running production environment is, by definition, a system with significant privileges inside that environment. doubleAI has not published a detailed security model for Maxwell’s early-access program, and that is a gap worth watching as more developers get hands-on access.
Why doubleAI Is Pushing a New Category Name
doubleAI’s decision to brand Maxwell as a Code Execution Model rather than simply calling it a compiler, an optimizer, or an AI agent is itself a deliberate marketing choice. Existing terms carry baggage. Calling Maxwell a compiler invites direct comparison with mature, well-understood tools like GCC and LLVM, which have decades of trust behind them and nowhere near the correctness concerns a newcomer would face. Calling it an AI agent lumps it in with code-generation tools, inviting questions about whether it writes new features rather than tuning existing ones. By coining a new term, doubleAI is trying to carve out a category where it can set the terms of comparison rather than compete on terms set by incumbents.
Whether that label sticks across the industry is an open question. Coverage of the launch from outlets including Globes has largely adopted doubleAI’s own terminology, which suggests the framing is landing with at least some of the press covering the Israeli tech scene. Terminology battles like this one tend to resolve only after a category has a handful of competing products, at which point a shared vocabulary becomes necessary for buyers trying to compare options.
What Comes Next
A handful of predictions follow naturally from where Maxwell stands today. First, expect doubleAI to publish more customer case studies modeled on the LTX example, since specific, named, quantified results are doing more to build credibility than the broader 38x and 4.5x figures on their own. Second, expect competitors to respond. A product category this directly tied to cloud cost savings will not stay uncontested for long, and infrastructure vendors with existing relationships inside large GPU fleets have an obvious incentive to build or acquire something similar.
Third, expect scrutiny of the formal verification claim to intensify as more developers get early access and start testing Maxwell against production-grade workloads rather than curated demos. Fourth, expect doubleAI’s pricing model, once it moves past early access, to be structured around a share of the compute savings it delivers rather than a flat subscription, since that is the pricing logic that lets a vendor make the strongest case to a finance team watching cloud spend. Fifth, and more speculatively, expect the term Code Execution Model to show up in other companies’ marketing within the next year, regardless of whether their underlying technology matches doubleAI’s approach, simply because the category name is catchy and the cost-savings pitch behind it is one every infrastructure buyer wants to hear right now.
Frequently Asked Questions
What is doubleAI’s Maxwell?
Maxwell is a product from the Israeli startup doubleAI that doubleAI calls a Code Execution Model. It observes a running program and rewrites the underlying machine code, or binary, to make it run faster on the specific hardware and data it encounters, without changing the program’s original source code.
Does Maxwell change my application’s source code?
No. doubleAI says Maxwell only changes the binary code running underneath an application. The original source code that developers wrote and maintain stays untouched.
How much faster does doubleAI claim Maxwell makes software run?
doubleAI reports more than 38 times faster performance on real-world data pipelines and more than 4.5 times faster performance on AI inference kernels. The company’s early-access page separately lists a 4.25 times figure for inference kernels, which differs slightly from the 4.5 times figure used in other company materials. These are company-reported numbers that have not been independently verified.
What hardware does Maxwell support?
doubleAI says Maxwell is designed to run on both CPUs and GPUs.
How is Maxwell different from a JIT compiler?
doubleAI co-founder Gal Beniamini noted that both approaches generate new machine code while a program runs. The company says Maxwell differs by working across runtimes and hardware types rather than inside one specific language runtime, and by formally proving each optimization preserves the original program’s behavior.
Who founded doubleAI?
doubleAI’s named co-founders are Amnon Shashua, also known as the co-founder and chief executive of Mobileye, and Shai Shalev-Shwartz.
Is Maxwell available to the public yet?
Maxwell is currently available through an early-access program for software developers. doubleAI has not announced general availability or public pricing.
What does doubleAI say about the formal correctness of Maxwell’s optimizations?
doubleAI says it formally proves that every optimization Maxwell applies is equivalent to the original code and preserves the program’s behavior. The company has not published detailed, independently verifiable proof methodology to date.
