doubleAI's Maxwell Automates GPU Kernel Optimization to Cut Enterprise Compute Bills

A Tel Aviv-based startup founded by two of the architects of Mobileye's perception stack has launched what it calls a new category of AI model — and quietly secured one of the largest Nvidia Blackwell hardware orders in the venture-backed ecosystem. doubleAI on Monday opened Maxwell, its first commercial product, for early access at heymaxwell.ai — a "code execution model" (CEM) that rewrites compiled software binaries to extract more performance from GPU and CPU hardware without requiring any changes to a customer's source code. The company disclosed $220M raised in total, anchored by a $200M Series A round co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with participation from Nvidia, Dell Technologies Capital, Greenfield, Pitango, BRM, and other investors. According to the Wall Street Journal, doubleAI spent roughly half that total acquiring 1,000 Nvidia B300 Blackwell Ultra GPUs to underpin a proprietary inference service.

A New Layer Between Model and Hardware

The AI inference stack has three conventional layers: the model itself, the serving framework, and the underlying hardware. Maxwell inserts a fourth. Rather than modifying a model's weights or swapping in new hardware, Maxwell operates at the compiled binary layer — inspecting how software actually executes on a chip and autonomously rewriting execution paths to use hardware resources more efficiently. Customers supply a compiled application; Maxwell returns an optimized version that runs faster on the same GPU or CPU without any source-code changes.

That distinction sidesteps the two most common bottlenecks in AI cost reduction: hardware swaps, which require capital and migration time, and manual kernel rewriting, which requires GPU engineers who can spend months tuning memory layout and thread scheduling for a single algorithm. Amnon Shashua, who co-founded the company alongside Mobileye collaborator Shai Shalev-Shwartz and four other researchers, frames doubleAI's mission as "Artificial Expert Intelligence" — narrow domain-specialist AI that surpasses human experts, contrasted with general-purpose AGI. "AI has changed how software is written, but not how it runs," Shashua said at the October launch.

The distinction from LLM-based coding assistants is deliberate. Tools like Claude Code, OpenAI Codex, and Gemini CLI generate source-level code that still requires a human to compile and test. doubleAI reported internally that such tools produced buggy solutions on roughly 40 percent of tasks in its own evaluations — a company-reported figure without disclosed methodology, but one consistent with known limitations of LLM-generated low-level code: GPU kernels interact with hardware in ways that are difficult for general-purpose models to reason about reliably.

How Maxwell Reasons About Correctness

The core technical challenge in automated kernel optimization is not generating a faster version of a kernel — it is verifying that the faster version is correct. GPU kernels are sensitive to subtle differences in memory alignment, warp divergence, and numerical precision; a rewritten kernel that passes superficial tests may fail on edge-case inputs or on a different GPU generation.

doubleAI addresses this with two proprietary methods it calls Diligent Learning and PAC Reasoning and Verification. Diligent Learning is a search process that explores the space of possible optimization decisions even when training data for a specific algorithm is sparse. PAC Reasoning borrows the "Probably Approximately Correct" framework from theoretical computer science — a formalism for guaranteeing that learned models generalize well under uncertainty — and applies it to the kernel correctness problem. Specifically, Maxwell generates adversarial test inputs designed to stress-test a rewritten kernel (Input Generation), then automatically judges whether outputs are correct without requiring human-labeled ground truth (Automatic Validation). According to the company's October launch announcement, each optimization is formally and mathematically proven to leave program behavior unchanged.

Read more: CoreWeave bond yield hits 14.5% as Wall Street reprices AI infrastructure risk

The Demo Trail: From Graph Library to Graph Theory

Before opening Maxwell for early access, doubleAI built a public record of what the underlying technology could accomplish. In March 2026, the company released an optimized version of Nvidia's cuGraph graph analytics library, rewriting every kernel in the library and reporting an average 3.6 times speedup over the expert-engineered originals. Every algorithm in the library ran faster; 55 percent of individual kernels ran more than twice as fast. The optimized library was released on GitHub as a drop-in replacement requiring no code changes — an architectural choice that validated the binary-compatibility claim.

The graph-library work was followed by demonstrations across different problem domains. On the pandas data manipulation library, doubleAI reported 38 times improvement in query performance on the PDS-H benchmark. For the LTX-2.5 video generation model, an early customer doubled inference speed. Across production AI inference workloads, the company reports kernels running more than 4.5 times faster. One unnamed enterprise customer reported a 67 percent reduction in compute spending — a figure cited by the Wall Street Journal but representing a single data point without disclosed baseline conditions.

Two demonstrations sit outside performance engineering but illuminate the PAC Verification framework. A doubleAI agent found a new algebraic construction for regular graphs that raised a girth coefficient unchanged since 1988 — the first improvement in 38 years on an open extremal graph theory problem — formalized in Lean for less than $2,000 in compute. Separately, Maxwell identified a chess configuration where 63 consecutive moves each give check, surpassing a decades-old record of 60. Neither demonstration is direct evidence that Maxwell generalizes to production LLM inference, but both show the system finding and verifying correct solutions in domains where ground truth is otherwise hard to establish algorithmically — the same underlying problem the PAC framework solves for kernel correctness.

A Hardware Bet Alongside the Software Business

The decision to spend approximately $110 million of $220 million raised on 1,000 Nvidia B300 Blackwell Ultra chips is the part of today's announcement that most clearly signals doubleAI's strategic intent. Shashua said the company cannot rely solely on renting compute as costs rise — SemiAnalysis data shows hourly H100 rental on a one-year contract has risen 60 percent over the past year. Owning a cluster of current-generation GPUs lets doubleAI run Maxwell's optimization workloads at scale and positions it to offer Maxwell-optimized inference as a service rather than merely selling software.

Read more: PilotDeck 2.0 cuts complex agent task costs by 83% using tiered model routing

That positioning echoes a pattern from Shashua's prior work. Mobileye became the dominant ADAS supplier by co-designing algorithms and hardware through the EyeQ chip family — a vertical integration that made competing on cost extremely difficult. doubleAI appears to be running the same playbook at the AI compute layer: own the software-hardware interface, validate on owned silicon, then offer the result as a managed service to enterprises that want efficiency without the operational complexity of running their own GPU clusters.

Where doubleAI Sits in the Optimization Stack

The tools for improving AI inference efficiency are numerous but each approaches the problem from a different angle. Nvidia's TensorRT-LLM compiles models with hardware-specific optimizations, but configuration is manual and output is hardware-locked. OpenAI's Triton lets developers write GPU kernels in Python but still requires expert programmers to express the optimization logic. vLLM and SGLang optimize request batching and KV-cache management for transformer inference without touching the underlying computation kernels. Intel's Xe-Forge project, described in a May 2026 preprint, is the closest research analog — LLM-guided kernel optimization for Intel GPUs — but targets different hardware and has no commercial availability.

Maxwell's differentiation is the combination of hardware-agnostic binary optimization, automated correctness verification, and a managed service. Shashua has named OpenAI and Anthropic as the class of customer that benefits most — organizations large enough that a 10 percent efficiency gain represents tens or hundreds of millions of dollars in annual savings — setting an ambitious opening target among customers who already employ expert GPU engineering teams.

What Comes Next for the Kernel Optimization Frontier

Maxwell's early access begins with demonstrated results on specific libraries — cuGraph, pandas, LTX video generation. The open question is whether the PAC Verification approach generalizes to the messier, less-structured landscape of production inference workloads for frontier language and multimodal models, where kernel behavior interacts with dynamic batching, KV-cache management, and speculative decoding in ways that narrow library optimization does not capture. The B300 chip purchase gives doubleAI a hardware platform to run those experiments at scale on the most capable available silicon.

If the approach generalizes, the market consequence is significant: enterprises and AI labs currently locked into a hardware upgrade cycle to cut compute costs would have a software-layer alternative. If it does not generalize cleanly beyond hand-selected library benchmarks, doubleAI will need to expand its demonstrated coverage substantially before targeting frontier model operators. The difference between those two outcomes will be visible in which customer names doubleAI is able to publish when Maxwell supply widens later this month.