Maxwell AI Cuts Code Costs 67%, Runs 38x Faster [2026]

Israeli startup doubleAI, co-founded by Mobileye and Open University founder Amnon Shashua, launched Maxwell on October 19, 2026, a "code execution model" that rewrites machine and binary code while a program is running, and is now in early access for developers. doubleAI claims Maxwell delivers more than 38x speedups on data pipelines, more than 4.5x gains on AI inference kernels, and a 67% compute cost reduction for one early customer, figures the company shared with reporters and that have not been independently audited. Maxwell leaves source code unchanged and works across CPUs and GPUs, positioning it as a runtime optimization layer rather than an IDE coding assistant.

An Israeli startup says it has built an AI system that can make software faster without a developer touching a single line of source code, and the claim is drawing attention across the software engineering world this week. doubleAI, co-founded by Mobileye and OpenU founder Amnon Shashua, introduced a product called Maxwell on October 19, 2026, describing it as a “code execution model” that rewrites machine and binary code while a program is actually running. The Register broke the story under the headline “Maxwell gives sub-optimal code a runtime glow-up,” and the framing has stuck: this is not another AI coding assistant that suggests edits in an IDE. It is software that tunes other software from the inside, after it has already shipped. Maxwell is now available to software developers through early access, and doubleAI is pitching it as the first entry in a new category it calls code execution models, or CEMs. The pitch is specific: more than 38x speedups on data pipelines, more than 4.5x gains on AI inference kernels, and a 67% compute cost reduction for one early customer, according to figures doubleAI shared with reporters. Those are doubleAI’s own numbers, not independently audited benchmarks, and that distinction matters for anyone deciding whether to trust the hype. Still, the idea behind Maxwell, that an AI system can continuously rewrite a program’s low-level instructions to match the hardware it happens to be running on, is new enough that it’s worth unpacking in detail. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What Is doubleAI’s Maxwell, and Why Is It Trending Today Maxwell is the flagship product from doubleAI, a company built around the idea that AI should optimize how code runs rather than how code is written. According to The Next Web https://thenextweb.com/news/doubleai-maxwell-code-execution-model-compute-costs , doubleAI describes Maxwell as a code execution model, a term the company is using to separate its product from large language models, coding copilots, and traditional compilers. The distinction is doing real work in the marketing: an LLM-based coding assistant edits source files before a human reviews and commits them. Maxwell, by contrast, operates after that point, inside the running binary, continuously inspecting instructions and swapping in faster equivalents as the program executes on real hardware with real data. That framing is why Maxwell jumped to the top of tech news feeds this week. Software performance engineering has historically been a scarce, expensive skill, the kind of work reserved for senior engineers tuning the hottest code paths at companies that can afford the headcount. A tool that claims to automate a meaningful slice of that work, and apply it continuously to code already in production, touches a nerve for anyone who manages cloud infrastructure bills or struggles to hire performance specialists. doubleAI is positioning Maxwell squarely at that pain point, and the timing lines up with a broader 2025-2026 trend of AI tools moving beyond code generation and into code execution. How Maxwell Rewrites Code Without Touching the Source The mechanics doubleAI has disclosed so far are relatively narrow but important. Maxwell runs alongside a program and rewrites the machine or binary code that the processor actually executes, while leaving the original source code completely unchanged. That means a development team’s codebase, version history, and review process stay untouched. Maxwell’s optimizations exist at a layer underneath all of that, closer to the hardware than to the code a developer would ever read or edit. Maxwell is built to work across both CPUs and GPUs, according to doubleAI’s own materials, which matters because performance bottlenecks in modern software rarely live in just one place. A data pipeline might be CPU-bound on serialization and I/O, while an AI inference workload is GPU-bound on kernel launches and memory bandwidth. A code execution model that only handled one processor type would miss half the opportunity. doubleAI’s early-access page states plainly that Maxwell’s goal is superoptimization, finding the fastest possible way to carry out a piece of code on the machine in front of it, rather than applying a fixed set of generic rules the way a conventional compiler does. Adapting to Hardware, Data, and Usage Patterns One detail that separates Maxwell’s pitch from a standard just-in-time compiler is adaptiveness. A venture capital firm backing the space, Lightspeed, described the category in a public statement: “The first CEM, doubleAI’s Maxwell, continuously inspects and rewrites code inside the runtime, fitting it to the hardware, data and usage it encounters,” Lightspeed said in a public post announcing the category. That phrase, fitting it to the hardware, data and usage it encounters, is the core claim: the same source code could end up compiled into different optimized machine code depending on whether it’s running on an older CPU generation, a specific GPU architecture, or a workload with unusual data distributions. Traditional compilers make their optimization decisions once, ahead of time, based on general assumptions about typical inputs. Just-in-time compilers, the kind used in Java virtual machines or modern JavaScript engines, do better by observing a program while it runs and recompiling hot code paths. The Register’s framing of Maxwell draws exactly that comparison: it says the system resembles a JIT compiler but claims more flexibility, continuously re-evaluating and re-optimizing based on live conditions rather than making a one-time decision after a warm-up period. The Performance Numbers: 38x, 4.5x, and a 67% Cost Cut The headline figures doubleAI is circulating are aggressive, and they deserve to be stated with their proper caveats. The company says Maxwell has made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster in real-world workloads. Separately, reporting from The Wall Street Journal, cited by The Next Web, says one early customer’s compute costs dropped by 67% after adopting Maxwell. doubleAI also points to LTX, a developer of open-weight multimodal AI models, which reportedly found that Maxwell doubled the speed of its model inference. None of these figures have been independently reproduced by a third party at the time of writing, and doubleAI has not published the benchmark hardware configurations, dataset sizes, or baseline comparisons that would let outside engineers verify the claims. That’s a familiar pattern in AI product launches: a headline multiplier gets the attention, and the methodology arrives later, if at all. The scale of the numbers, especially the 38x figure, is the kind of result that invites scrutiny rather than blind acceptance, and engineers evaluating Maxwell during its early access period should expect to run their own workloads before trusting any of these figures for budget planning. | Metric | Reported Result | Source | |---|---|---| | Data pipeline speed | More than 38x faster | doubleAI, via The Next Web | | AI inference kernel speed | More than 4.5x faster | doubleAI, via The Next Web | | Early customer compute cost | Reduced by 67% | Wall Street Journal, via The Next Web | | LTX model inference speed | 2x faster | doubleAI, via The Next Web | | Supported hardware | CPUs and GPUs | doubleAI early access page | | Availability | Early access for developers | doubleAI, The Register | Formal Verification: doubleAI’s Answer to “What If It Breaks My Code” The obvious objection to any system that automatically rewrites machine code while a program runs is correctness. If an AI model is swapping out instructions on the fly, what stops it from introducing a subtle bug that only shows up under rare conditions, long after a human reviewer signed off on the original source? doubleAI’s answer is formal verification: the company says every optimization Maxwell applies is formally proven equivalent to the original code before it takes effect, not just tested or sampled, but mathematically verified to produce the same behavior. This is the part of Maxwell’s pitch that separates it most clearly from a typical AI coding tool, where hallucinated or subtly wrong suggestions are an accepted cost that human review is supposed to catch. Formal verification, when done correctly, offers a mathematical guarantee rather than a statistical one. doubleAI frames this as the central engineering challenge the company had to solve to make a continuously rewriting runtime system viable at all: speed gains are worthless to an engineering team if they come with any risk of silently incorrect output. On its own early-access page, doubleAI describes the product’s ambition this way: “Its goal is superoptimisation: finding the fastest possible way to carry out a piece of code on the machine in front of it,” a framing that places correctness as the non-negotiable floor under which all the speed claims sit, according to the heymaxwell.ai early access page https://heymaxwell.ai/early-access . Whether that verification claim holds up under adversarial scrutiny, from academic formal-methods researchers or from large engineering organizations running Maxwell at scale, is one of the more important open questions hanging over this launch. Formal equivalence proofs are notoriously hard to generate automatically for arbitrary real-world code, especially code involving concurrency, floating-point arithmetic, or hardware-specific instructions, which happen to be exactly the kinds of code where performance optimization matters most. Inside doubleAI: Amnon Shashua’s New Venture doubleAI’s profile is boosted considerably by its founder. Amnon Shashua is named in coverage as a co-founder of the company, and he carries a track record that gives Maxwell more immediate credibility than a typical early-access product launch would get on its own. Shashua previously co-founded Mobileye, the autonomous-driving technology company that Intel acquired, and he has a long academic background in computer vision and machine learning. A profile of the launch from Israeli business outlet Globes https://en.globes.co.il/en/article-amnon-shashuas-doubleai-unveils-code-execution-model-1001560057 frames doubleAI’s Maxwell launch as a notable bet from a founder with a history of building technically ambitious companies that eventually found large acquirers or large markets. That history matters for how the market is likely to read this launch. A code execution model is a technically daring concept, automatically rewriting running binaries with formally verified equivalence is not a trivial engineering problem, and skepticism would be the default reaction if the claim came from an unknown startup with no track record. Shashua’s involvement shifts the conversation from “is this even possible” toward “how well does the current version actually perform,” which is a meaningfully different starting point for an early-access launch. Maxwell vs Just-in-Time Compilers: What’s Actually New It’s worth being precise about what Maxwell claims to add on top of decades of existing compiler and runtime research, because “AI makes code faster” is not, by itself, a new idea. Just-in-time compilers have been rewriting bytecode into optimized machine code at runtime since the 1990s. Profile-guided optimization has let compilers use real execution data to inform later compilation passes for almost as long. Superoptimizers, tools that search exhaustively or heuristically for the fastest possible instruction sequence for a small code fragment, are an academic research area going back to the 1980s. What doubleAI is claiming is a combination of three things that have rarely, if ever, shipped together in a production tool: continuous re-optimization that doesn’t stop after an initial warm-up period, a search space wide enough to span both CPU and GPU instruction sets, and formally verified correctness rather than testing-based confidence. The Register’s comparison to JIT compilers captures the lineage accurately, Maxwell is clearly a descendant of that research tradition, but the company’s claim is that it operates with more flexibility and a stronger correctness guarantee than anything currently deployed at scale. Readers tracking the broader shift toward AI-assisted developer tooling, including AI-driven code review systems like Cursor’s Bugbot https://shattered.io/set-up-cursor-bugbot-2026/ , will recognize the pattern: AI is steadily moving from suggesting code to writing code to, now, continuously re-optimizing code after it’s already running. Who’s Using Maxwell So Far Beyond the unnamed early customer cited by the Wall Street Journal, the one named early adopter in current coverage is LTX, an open-weight multimodal AI model developer. doubleAI says LTX found that Maxwell doubled its model inference speed, a figure that, if it holds up under broader deployment, would be a meaningful cost reduction for any company running inference at scale. Doubling inference speed on the same hardware is roughly equivalent to halving the GPU fleet needed to serve the same traffic, which is the kind of number that gets a CFO’s attention as quickly as a CTO’s. Maxwell’s current status as an early-access product for software developers, rather than a generally available commercial release, means the customer list is necessarily short right now. doubleAI has not disclosed pricing, a public release timeline, or the specific criteria developers need to meet to get into the program. That lack of detail is normal for a product at this stage, but it also means most of what’s circulating about Maxwell right now is doubleAI’s own framing of its own results, repeated across outlets that covered the same announcement. The Code Execution Model Category: A New Label for an Old Problem doubleAI’s decision to coin a new category name, code execution model, rather than describing Maxwell as “an AI compiler” or “an AI performance tool,” is a deliberate positioning choice. Category names matter in enterprise software sales: a company that successfully defines a new category gets to set the terms other competitors are measured against, the way Datadog did with observability or Snowflake did with the data cloud. Whether CEM becomes an industry-standard term or fades as marketing language will depend heavily on whether other vendors adopt the framing, and whether Maxwell’s real-world results hold up once more engineering teams get access. What’s notable is that the category sits adjacent to, but distinct from, several trends already reshaping how software gets built and run. AI coding assistants have already normalized the idea of AI touching production code. Rising interest in developer tooling that automates code quality, like the kind of changes covered in recent updates to Chrome DevTools’ AI agent controls https://shattered.io/chrome-devtools-154-ai-agent-controls-2026/ , shows browser and IDE vendors racing to give AI systems more autonomy over developer workflows. Maxwell extends that autonomy one layer further down, into the compiled binary itself. Historical Context: From Manual Tuning to Autonomous Runtime AI Performance engineering has moved through several distinct eras. In the 1990s and 2000s, tuning meant a small number of specialized engineers hand-optimizing assembly or rewriting hot loops after profiling with tools like gprof or VTune. The 2010s brought profile-guided optimization and more sophisticated JIT compilers into mainstream use, reducing but not eliminating the need for manual tuning. The late 2010s and early 2020s saw the rise of auto-tuning frameworks for specific domains, like TVM and Halide for machine learning kernels, which used search and heuristics to generate optimized code for particular hardware targets, but typically as an offline compilation step rather than a continuous runtime process. Maxwell’s pitch represents a further step: taking that auto-tuning concept and running it continuously, in production, across both general-purpose CPU workloads and specialized AI inference kernels, backed by formal verification instead of empirical testing. It’s a natural, if ambitious, extension of a multi-decade trend toward automating work that used to require scarce human expertise. The same trend is visible in the broader AI chip and infrastructure market, where hardware makers are racing to build silicon that can keep pace with AI workloads, a theme that showed up recently in coverage of Arm’s AGI CPU claims against AMD’s EPYC lineup https://shattered.io/arm-agi-cpu-beats-amd-epyc-9755-1-99x-rack-2026/ and in Intel’s push into large-memory AI accelerators aimed at Nvidia and AMD https://shattered.io/intel-480gb-gpu-d-series-vs-nvidia-amd-2026/ . Software-level optimization tools like Maxwell and hardware-level investments are two sides of the same cost-reduction pressure facing every company running large-scale AI workloads in 2026. Market Impact: What This Means for Cloud Compute Spending If Maxwell’s reported numbers generalize even partially beyond doubleAI’s cited examples, the market implications reach well past one startup’s product launch. Cloud compute spending has become one of the largest line items for AI-heavy companies, and inference costs in particular have drawn intense scrutiny as AI products scale from pilot projects to production traffic. A tool that can meaningfully cut inference costs without requiring a rewrite of application code is attractive precisely because it sidesteps the biggest obstacle to adopting most performance improvements: engineering time. Rewriting a data pipeline or inference stack by hand to capture a 38x or 4.5x speedup, assuming that were even achievable manually, could take a specialized team months of work, with real risk of introducing bugs along the way. A tool that claims to deliver comparable gains automatically, with a formal correctness guarantee and no source code changes, removes both the time cost and much of the risk, at least on paper. That’s the value proposition doubleAI is selling to prospective enterprise customers, and it’s also exactly the kind of claim that will get tested hardest once more companies than LTX and doubleAI’s unnamed early customer start running Maxwell against production workloads with public benchmarks attached. A Pattern Playing Out Across the AI Infrastructure Stack Cost-cutting AI infrastructure announcements have been a recurring theme through 2025 and 2026, as the industry works through the economics of running increasingly large models at increasingly large scale. Memory and framework-level efficiency gains, like the kind described in recent coverage of Zod 4.6’s memory footprint reduction alongside React 19.3.1 https://shattered.io/react-19-3-zod-4-6-memory-performance-2026/ , show the same underlying pressure playing out at a different layer of the stack: developers and infrastructure teams are hunting for efficiency wherever they can find it, whether that’s a smaller JavaScript validation library or a runtime system rewriting compiled binaries. Maxwell is simply the most aggressive version of that hunt disclosed so far this year. Competitive Landscape: Where Maxwell Fits Among AI Coding Tools Maxwell doesn’t compete directly with the current wave of AI coding assistants, and that’s worth spelling out clearly, because the categories are easy to conflate. Tools like GitHub Copilot https://shattered.io/github-copilot-retires-6-models-gpt-5-6-2026/ , Cursor, and similar AI pair-programming products operate at write time, suggesting or generating source code that a human developer reviews, edits, and commits. Maxwell operates at run time, after code has already been committed, built, and deployed, working on the compiled artifact rather than the source. In principle, a team could use both: an AI coding assistant to help write the application, and Maxwell running underneath it in production to optimize the compiled result. | Approach | When It Operates | What It Modifies | Correctness Approach | Target Hardware | |---|---|---|---|---| | Maxwell CEM | During program execution | Machine/binary code | Formal equivalence verification doubleAI claim | CPUs and GPUs | | Traditional JIT compiler | During execution, rule-based | Bytecode to machine code | Deterministic compilation rules | Primarily CPU | | AI coding assistants | Before execution, at write time | Source code suggestions | Human developer review | Not applicable | | Manual performance engineering | Ongoing, human-driven | Source code and build configuration | Human testing and benchmarking | Varies by project | This complementary positioning is smart from a go-to-market perspective. doubleAI isn’t asking engineering teams to abandon their existing toolchain or workflow, it’s asking them to add a runtime layer underneath whatever they’re already running. That’s a much easier sales pitch than demanding teams change how they write code, and it mirrors how observability and security tools have successfully sold into enterprises for years: slot in alongside the existing stack, prove value quickly, expand from there. The Engineering Challenge Maxwell Is Trying to Solve To understand why doubleAI frames correctness verification as such a central part of its pitch, it helps to think conceptually about what continuous runtime rewriting actually requires. A simplified illustration of the kind of guarantee a system like this needs to provide looks something like this, where an optimization is only applied once it has been checked against the original behavior: // Conceptual illustration only -- not an actual doubleAI code sample if formally verify equivalent original instructions, candidate instructions { runtime.replace original instructions, candidate instructions ; } else { runtime.keep original instructions ; } The hard part, and the reason this hasn’t been commercialized widely before now, is that proving equivalence automatically for arbitrary real-world code is a genuinely difficult formal-methods problem, especially once concurrency, floating-point rounding behavior, and hardware-specific instructions enter the picture. doubleAI says Maxwell is doubleAI’s artificial expert intelligence for performance engineering, built specifically to design, verify, specialize, and tune code at that level, a description that frames the formal verification step as inseparable from the speed claims rather than a bolt-on safety feature. Risks and Open Questions Facing Maxwell Several questions remain genuinely unresolved as Maxwell moves through early access. First, independent verification: no third-party benchmark of Maxwell’s claimed speedups has surfaced yet, and the 38x and 4.5x figures both come from doubleAI’s own disclosures. Second, scope: it’s unclear from current reporting how broad a range of code Maxwell can meaningfully optimize, versus how narrow the specific workloads were that produced its best-case numbers. Third, security and auditability: a system that continuously rewrites running binaries introduces a new layer that security and compliance teams will need to understand, especially in regulated industries where auditors expect to trace exactly what code is executing in production at any given time. There’s also the commercial question of pricing and availability, neither of which doubleAI has disclosed publicly. Early access programs for ambitious infrastructure products often take months to mature into generally available releases with transparent pricing, and plenty of promising runtime and compiler technologies over the years have struggled to cross that gap from impressive demo to dependable production tool used widely across an industry. Predictions: Where Runtime Code Optimization Goes From Here - Expect at least one major cloud provider or hyperscaler to publicly evaluate or partner with doubleAI within the next two to three quarters, given how directly Maxwell’s cost claims speak to cloud infrastructure margins. - Independent benchmarking firms and academic formal-methods researchers are likely to scrutinize doubleAI’s verification claims once Maxwell reaches a wider pool of early-access developers, and the 38x figure in particular will face pressure to be reproduced on standardized workloads. - Competing code execution model products are likely to emerge from other AI infrastructure startups and possibly from established compiler teams at major chipmakers, now that doubleAI has demonstrated investor and press appetite for the category. - AI coding assistant vendors will likely begin marketing deeper runtime integration or partnerships, positioning write-time AI tools and run-time tools like Maxwell as complementary parts of a single AI-native development pipeline. - Enterprise adoption will probably move slowly at first, concentrated among AI-heavy companies with large inference bills, before if results hold up spreading to more general backend and data infrastructure workloads. What Developers Should Watch Next For engineering teams curious about Maxwell, the practical next step is simply watching for independent, reproducible benchmarks rather than acting on doubleAI’s own figures alone. Early access programs are, by design, curated to showcase best-case results, and the gap between a vendor’s best-case demo and a typical production outcome is often the single most important number a buyer never sees up front. Teams already experimenting with AI across their developer pipeline, from code review automation to AI-assisted infrastructure decisions discussed in recent coverage of the rapid pace of new AI model releases this fall https://shattered.io/what-changed-in-october-2026-two-new-model-families-one-week-apart/ , should treat Maxwell the same way they’d treat any new performance tool: pilot it against a representative workload, measure independently, and only then decide whether the formal verification guarantee and the reported speedups justify adopting a new layer in the production stack. Frequently Asked Questions What is doubleAI’s Maxwell? Maxwell is a code execution model from doubleAI, a company co-founded by Amnon Shashua. It rewrites machine or binary code while a program is running, without changing the program’s original source code, with the stated goal of making existing software faster on the CPUs and GPUs it actually runs on. Is Maxwell the same thing as an AI coding assistant like Copilot or Cursor? No. AI coding assistants generate or suggest source code before a human review and commit. Maxwell operates after that point, optimizing already-compiled binary code at runtime, which means it works underneath, not instead of, tools like Copilot or Cursor. How fast does Maxwell claim to make code? doubleAI says Maxwell has made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster in real-world workloads. One early customer reportedly cut compute costs by 67%, and LTX reportedly saw a 2x speedup in model inference. These are doubleAI’s own disclosed figures and have not yet been independently verified. How does Maxwell avoid breaking the programs it optimizes? doubleAI says every optimization Maxwell applies is formally proven equivalent to the original code before it takes effect, which is a mathematical correctness guarantee rather than a testing-based one. The strength of this claim has not yet been independently examined by outside formal-methods researchers. What hardware does Maxwell support? According to doubleAI, Maxwell is designed to work across both CPUs and GPUs, which allows it to target bottlenecks in general-purpose workloads as well as specialized AI inference kernels. Who is behind doubleAI? doubleAI was co-founded by Amnon Shashua, previously a co-founder of Mobileye. His track record in building technically ambitious companies has drawn added attention to Maxwell’s launch, beyond what a typical early-stage product announcement would receive. Can I use Maxwell today? Maxwell is currently available through early access for software developers. doubleAI has not publicly disclosed pricing or a timeline for general availability. What is a “code execution model,” and is it a real industry category yet? Code execution model, or CEM, is a term doubleAI is using to describe Maxwell’s category: AI that optimizes compiled code during execution rather than generating source code. It’s a new label coined around this launch, and whether it becomes a widely adopted industry category will depend on whether other vendors build comparable products and whether Maxwell’s results hold up at broader scale.