doubleAI
AI infrastructure startup optimizing software execution to reduce compute requirements without rewriting code.
Website: https://www.doubleai.com
Cover Block
Publicly reported
| Field | Detail |
|---|---|
| Name | doubleAI |
| Tagline | AI infrastructure startup optimizing software execution to reduce compute requirements without rewriting code [Calcalist Tech, October 2026] |
| Headquarters | Tel Aviv, Israel |
| Founded | 2023 |
| Stage | Series A |
| Business model | B2B |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | Middle East / North Africa |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (3+) |
| Funding label | $100M+ |
| Total disclosed funding | $220,000,000 [The Wall Street Journal, October 2026] |
Links
Publicly reported
- Website: https://www.doubleai.com
- Careers: https://www.doubleai.com/careers
- DevOps Engineer: https://www.doubleai.com/careers/devops-engineer
- Product Engineer: https://www.doubleai.com/careers/product-engineer
- Deep Learning Engineer: https://www.doubleai.com/careers/deep-learning-engineer
- Research Engineer: https://www.doubleai.com/careers/research-engineer
Summary and Signal
PUBLIC doubleAI is building AI infrastructure that aims to cut compute usage at the execution layer, a timely pitch as model training and inference costs remain a central constraint for enterprise AI adoption [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026]. Founded in 2023 in Tel Aviv, the company emerged with unusual financial backing for its age, reaching $220 million in disclosed funding by October 2026 and attracting a syndicate that includes Lightspeed Venture Partners, Bessemer Venture Partners, NVIDIA, Dell Technologies Capital, Pitango, BRM Group, and Greenfield [The Wall Street Journal, October 2026] [Calcalistech, October 2026].
Its first product, Maxwell, is described as a system that analyzes running programs, hardware, and workloads, then optimizes machine instructions without requiring changes to application source code, which places the differentiation below the model and application layers where most AI tooling companies are currently competing [Calcalist Tech, October 2026] [The Next Web, October 2026]. Early traction claims are promising but still narrow in public evidence: doubleAI has said one early customer cut compute costs by 67%, and LTX reported that Maxwell doubled model-inference speed in its environment [The Next Web, October 2026].
The founding story matters here because the team is not starting from a cold market entry. Co-founder and CEO Amnon Shashua previously led Mobileye, while co-founder and CTO Shai Shalev-Shwartz also held senior leadership responsibilities there; public reporting and LinkedIn-sourced references also tie the broader founding bench to deep technical research and systems work [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026] [The Next Web, October 2026].
Commercially, the notable point is not only the scale of financing but the proposed pricing logic. According to Calcalist Tech and additional cited reporting, doubleAI plans to charge customers based on a share of the compute savings it generates, which could align spend to realized value if the savings prove durable across heterogeneous workloads [Calcalist Tech, October 2026] [재경일보].
Over the next 12 to 18 months, the main questions are straightforward. Investors should watch for broader proof that Maxwell's gains generalize beyond a small set of early examples, evidence that savings-based pricing converts into repeatable enterprise contracts, and whether the company can translate an elite technical founding team and roughly 40-person organization into a defensible go-to-market motion before larger infrastructure vendors push deeper into runtime optimization [Calcalist Tech, October 2026] [The Next Web, October 2026].
One source, partially checked -- This section relies on strong corroboration for founding, funding, and headquarters, but material product and performance claims remain primarily company-linked or single-source reported.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series A |
| Business Model | B2B |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | Middle East / North Africa |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (3+) |
| Funding | $100M+ (total disclosed ~$220,000,000) |
Company Overview
PUBLIC
The public record on doubleAI is still thin, but the core outline is clear. The company was founded in 2023 and is headquartered in Tel Aviv, Israel, according to its careers materials and company profile references [doubleAI, October 2026].
The chronology that can be stated from the allowed source set is narrow. By October 2026, the company was publicly recruiting for infrastructure, product, deep learning, and research engineering roles, which is consistent with an operating company moving from research into broader productization [doubleAI, October 2026]. The same careers materials describe a mission around models, agents, and verification methods for AI systems, providing the clearest company-issued description of its direction at this stage [doubleAI, October 2026].
On legal entity details, the allowed public sources for this section do not establish more than headquarters and operating identity. Investors should treat the company as an Israel-based startup with a 2023 founding date until state filings or registry extracts are surfaced directly [doubleAI, October 2026].
No independent source found -- This section relies primarily on company website materials, with no Crunchbase or state filing evidence supplied in the source set.
The Product and the Stack
MIXED
The technical claim worth isolating is straightforward: doubleAI says it is working at the execution layer rather than the code-generation layer. Public reporting describes its first product, Maxwell, as a system that analyzes a program while it runs, alongside the underlying hardware and workload, then optimizes machine instructions to reduce compute requirements without changing the application’s source code [Calcalist Tech, October 2026] [The Next Web, October 2026]. Company materials make the claim more explicit, describing Maxwell as software that observes a running program, rewrites the binary underneath for the specific machine and data, and applies formal verification to check equivalence with the original behavior [doubleAI, October 2026]. Those are company and press descriptions rather than independent benchmark studies, so the capability should be read as reported product behavior, not yet broadly validated market fact.
The early performance evidence is promising but still narrow. The Next Web reported that one early customer cut compute costs by 67%, and that LTX said Maxwell doubled model-inference speed [The Next Web, October 2026]. Company-linked materials add more granular examples, including reported gains in data pipelines, inference kernels, and a reduction in LTX 4K video generation time from 250 seconds to 134 seconds per clip [doubleAI, October 2026]. Even taken at face value, these examples suggest the current wedge is expensive CPU and GPU workloads where marginal efficiency gains have direct budget impact, and the reported pricing model reflects that logic, with doubleAI planning to charge based on a share of compute savings generated for customers [Calcalist Tech, October 2026] [재경일보].
The public hiring footprint points to a company still building across research, infrastructure, and productization. As of October 2026, the careers page listed openings for DevOps Engineer, Product Engineer, Deep Learning Engineer, and Research Engineer, with role descriptions touching infrastructure, training and inference systems, and reasoning or code-optimization agents [doubleAI, October 2026]. The broader stack is therefore best described conservatively as applied systems research plus production AI infrastructure, with any deeper stack description remaining an inference from job postings rather than a confirmed architecture map [doubleAI, October 2026].
No independent source found -- Core product behavior and benchmark claims rely materially on company materials and press accounts that cite company-reported results, with only partial corroboration from independent coverage.
The Market They Are Entering
PUBLIC
The market matters now because the cost of running modern AI and data workloads has become a first-order constraint for buyers, while doubleAI is positioned around reducing that spend rather than asking customers to rebuild applications from scratch [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026].
The public record here is thinner than investors would usually want for a formal TAM exercise. No named third-party market report is captured in the source set for runtime optimization software, code execution models, or binary-level AI infrastructure, so any hard market sizing would overstate the evidence. What can be said from reporting is that doubleAI is targeting organizations with expensive CPU and GPU workloads, and it presents Maxwell as infrastructure that improves software performance while preserving application source code, which places the company at the intersection of AI infrastructure, performance engineering, and cost-optimization tooling [Calcalist Tech, October 2026] [The Next Web, October 2026].
Demand drivers are more visible than market size. The company launched Maxwell into a backdrop where model training and inference costs remain material enough that a product promising lower compute requirements can earn attention quickly, particularly if the deployment does not require codebase rewrites [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026]. Early proof points cited in public coverage point in the same direction, although they remain company-linked and should be read cautiously: one early customer reportedly cut compute costs by 67%, and LTX reported that Maxwell doubled model-inference speed [The Next Web, October 2026].
Adjacent markets are easier to identify than direct category boundaries. Maxwell is described as operating below the application layer and above hardware, which makes the nearest substitutes a mix of compiler tooling, inference optimization stacks, GPU kernel tuning, and broader cloud cost-management software rather than coding copilots or model builders [Calcalist Tech, October 2026]. The careers page also broadens the frame by describing the company as working on models, agents, and verification methods for AI systems that reason deeply and learn from experience, suggesting a longer-term ambition that could touch both systems software and agentic AI infrastructure if execution holds [doubleAI, October 2026].
Macro and regulatory forces cut in both directions. On the supportive side, persistent pressure to make AI workloads cheaper should help products that improve utilization on existing CPUs and GPUs, especially for teams already constrained by infrastructure budgets or hardware access [The Wall Street Journal, October 2026] [The Next Web, October 2026]. On the limiting side, there is no cited public evidence yet that this optimization layer is becoming a standard budget line item across enterprises, and buyers in sensitive environments may require unusually strong validation around correctness, reproducibility, and system behavior before adopting software that rewrites execution paths underneath production workloads, even when source code is preserved [Calcalist Tech, October 2026] [heymaxwell.ai].
| Market frame | Public evidence | Implication |
|---|---|---|
| AI infrastructure cost optimization | Maxwell is positioned to reduce compute requirements for running software on CPUs and GPUs without changing source code [Calcalist Tech, October 2026] | Points to demand from teams with high recurring compute spend rather than from general developer tooling budgets |
| Inference performance tooling | LTX reported that Maxwell doubled model-inference speed [The Next Web, October 2026] | Suggests early relevance for model-serving and media-generation workloads, if replicated beyond design partners |
| Broader systems and agent infrastructure | The company describes its mission as building models, agents, and verification methods for AI systems that reason deeply and learn from experience [doubleAI, October 2026] | Indicates a wider ambition than one optimization product, but that expansion is still strategic positioning rather than demonstrated market presence |
The table shows a market thesis built more on acute customer pain than on published category sizing. That can still support a large outcome, but at this stage the addressable market is better framed as a convergence of existing infrastructure budgets than as a clean, externally validated software category.
No independent source found -- This section relies on named media coverage for positioning and demand context, but it has no captured third-party market report for TAM/SAM/SOM and uses company-linked product framing for part of the market definition.
The Competitive Field
Competitive map
Public record plus analysis The company is positioning itself in a less crowded layer of the stack than most AI infrastructure startups: not model training, not code generation, but execution-time optimization that sits between application code and hardware, with Maxwell described as preserving source code while rewriting machine instructions as software runs [Calcalist Tech, October 2026] [The Next Web, October 2026].
On the public record, that means the relevant comparison set breaks into three groups rather than one direct peer set. First are incumbent hardware and systems vendors, especially NVIDIA, whose software stack and baseline kernels define much of the performance starting point on modern GPU workloads; doubleAI's own materials frame Maxwell against PyTorch baselines and GPU kernel optimization work, which implies the practical alternative for many teams is to keep tuning inside the vendor stack rather than adopt a new optimization layer [LinkedIn, Cristian-George Farauanu, retrieved 2026] [LinkedIn, Benjamin Guggenheim, retrieved 2026]. Second are model and application teams such as LTX, which can pursue performance through internal engineering, compiler work, and handcrafted kernels; LTX's reported near 2x speedup with Maxwell matters because it suggests even already-optimized workloads may still have room for a third-party optimizer [The Next Web, October 2026] [doubleAI, retrieved 2026]. Third are adjacent AI developer tools, including code assistants and model platforms, that compete for infrastructure budgets but do not appear to attack the same bottleneck directly [Calcalist Tech, October 2026].
The defensible edge today looks less like distribution and more like concentration of technical talent and capital. Amnon Shashua's record at Mobileye gives the company immediate credibility with investors and technical recruits, while the investor list, including NVIDIA, Dell Technologies Capital, Lightspeed Venture Partners, Bessemer Venture Partners, BRM Group, Pitango, and Greenfield, reduces financing risk for a category that likely needs long development cycles and heavy compute access [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026]. That edge is real, but it is also perishable: talent and capital help a company establish an execution layer, yet they do not by themselves prevent a platform owner or hyperscaler from building similar optimization features into the default toolchain if the category proves commercially attractive [Calcalist Tech, October 2026].
The main exposure is to vertical integration by incumbents rather than to a named startup rival in the current source set. NVIDIA is both a strategic backer and the most obvious reference point because the company already owns critical parts of the AI software and hardware stack, and Maxwell's reported outperformance on selected Blackwell kernels only underscores that doubleAI is competing near terrain the platform vendor can choose to defend aggressively [LinkedIn, Cristian-George Farauanu, retrieved 2026] [The Wall Street Journal, October 2026]. There is a second exposure on proof: reported gains such as a 67% compute cost reduction for one early customer and doubled inference speed for LTX are promising, but they remain customer-specific and largely company-mediated evidence rather than a broad, independently benchmarked moat across sectors [The Next Web, October 2026].
Over the next 18 months, the most plausible competitive scenario is that the category splits by workload complexity and willingness to change production infrastructure. NVIDIA is the likely winner if customers decide vendor-native tooling is sufficient and prefer a single-stack procurement path, particularly for standard training and inference pipelines [The Wall Street Journal, October 2026]. doubleAI is the likely winner if Maxwell keeps showing material gains on already-optimized CPU and GPU workloads without forcing source-code changes, because that would define a narrow but valuable control point in the compute stack [Calcalist Tech, October 2026] [The Next Web, October 2026]. The likely loser if that proof fails is not a named startup in the available sources, but the standalone execution-optimization category itself, since buyers can fall back to internal performance engineering or incumbent tooling without much category-switching cost [Calcalist Tech, October 2026].
No independent source found -- This section relies on named public reporting for company positioning and investor base, but the core competitive claims about technical differentiation and benchmark outperformance are primarily company statements, company-linked research posts, or employee LinkedIn posts [Calcalist Tech, October 2026] [The Next Web, October 2026] [doubleAI, retrieved 2026] [LinkedIn, Cristian-George Farauanu, retrieved 2026].
Opportunity
PUBLIC
The prize here is unusually large if doubleAI can prove that execution-time optimization is repeatable across mainstream CPU and GPU workloads, because compute cost has become a board-level constraint for AI teams and infrastructure owners rather than a narrow developer pain point [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026].
In plain language, the biggest plausible outcome is that doubleAI becomes a default optimization layer that sits between expensive software workloads and the hardware they run on. That outcome is reachable, not merely aspirational, because the company is not entering with a generic productivity story. It is entering with a product, Maxwell, that is described as optimizing software as it runs on CPUs and GPUs without requiring source-code changes, a pricing model tied to realized compute savings, and early public examples of measurable performance gains from launch coverage and company materials [Calcalist Tech, October 2026] [The Next Web, October 2026] [doubleAI, October 2026]. The additional signal is capital formation. A startup founded in 2023 that had raised $220 million by October 2026 from investors including NVIDIA, Dell Technologies Capital, Lightspeed Venture Partners, and Bessemer Venture Partners has enough backing to pursue a technically demanding platform wedge rather than a narrow consulting motion [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026].
The upside can be framed through a small number of concrete paths to scale.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Enterprise compute governor | Maxwell becomes a standard layer for enterprises running costly inference, analytics, and data workloads across mixed CPU and GPU environments | A public customer set shows repeatable savings beyond a single early deployment, validating savings-based pricing [The Next Web, October 2026] [Calcalist Tech, October 2026] | Maxwell is positioned to optimize running programs without rewriting source code, which lowers adoption friction relative to a full application rebuild [Calcalist Tech, October 2026] |
| AI infrastructure standard for model teams | doubleAI wins adoption among AI-native companies that already tune inference pipelines heavily, then expands from model inference into adjacent performance bottlenecks | A technically sophisticated customer such as LTX continues to report material gains, showing the product still matters even after internal optimization [The Next Web, October 2026] [doubleAI, October 2026] | Public reporting says LTX saw roughly 2x inference-speed improvement, and company research argues those gains came on top of an already optimized stack [The Next Web, October 2026] [doubleAI, October 2026] |
| New software abstraction layer | The company broadens from Maxwell into a family of code-execution and verification systems that developers treat as infrastructure, not a point tool | The broader platform thesis, models, agents, and verification methods, turns into adjacent products for optimization and assurance [doubleAI, October 2026] | The careers page describes a mission beyond one optimizer, and current hiring across research, product, deep learning, and DevOps suggests a platform buildout rather than a single-feature company [doubleAI, October 2026] |
What compounding could look like is straightforward, even if the evidence is still early. If doubleAI can attach pricing to realized compute savings, each successful deployment should do three things at once: create a clear budgetary ROI case, generate more workload coverage for the optimization engine, and make future sales easier because the product does not require source-code rewrites to show value [Calcalist Tech, October 2026]. The first signs of that flywheel are visible in public materials, though still company-led. Maxwell is framed as working across both CPUs and GPUs, and public posts tied to the company point to gains in data pipelines, inference kernels, and video generation workflows, which suggests a path from one high-value use case into a broader performance layer if those results generalize [Calcalist Tech, October 2026] [doubleAI, October 2026] [LinkedIn, Cristian-George Farauanu].
The size of the win is best expressed as a scenario rather than a forecast. The company was reported at approximately a $1 billion valuation around its November 2025 Series A, with more than $220 million raised by October 2026 [Calcalistech, October 2026] [The Wall Street Journal, October 2026]. If the strongest scenario plays out, meaning doubleAI proves it can become the standard software layer for reducing compute intensity across enterprise and AI-native workloads, the outcome would likely be many multiples above that level because it would be selling into a cost base measured in ongoing infrastructure spend rather than a one-time tooling budget (scenario, not a forecast). Public sources provided for this report do not include a named market-size study or a directly comparable public peer with the same execution-layer positioning, so the upside case rests on the economics of compute reduction and the unusually strong investor and founder base rather than a neat TAM model [The Wall Street Journal, October 2026] [Calcalist Tech, October 2026].
No independent source found -- This section relies on a mix of independent reporting for funding, team, and product framing, but the core upside case depends materially on company-reported product claims and early customer outcomes that remain lightly corroborated.
Sources
Publicly reported
[The Wall Street Journal, October 2026] Nvidia-Backed Startup doubleAI Launches Model to Pare Cost of Compute | https://www.wsj.com/cio-journal/nvidia-backed-startup-doubleai-launches-model-to-pare-cost-of-compute-b780f215
[Calcalist Tech, October 2026] Shashua’s doubleAI unveils AI models that optimize … | https://www.calcalistech.com/ctechnews/article/sk11xpc73ge
[The Next Web, October 2026] doubleAI’s Maxwell rewrites code as it runs to cut compute costs | https://www.thenextweb.com/news/doubleai-maxwell-code-execution-model-compute-costs
[doubleAI, October 2026] Careers | https://www.doubleai.com/careers
[TechCrunch, 2026-07-23] Mobileye CEO Amnon Shashua to step aside as company pushes into robotaxis, robotics | https://techcrunch.com/2026/07/23/mobileye-ceo-amnon-shashua-to-step-aside-as-company-pushes-into-robotaxis-robotics/
[LinkedIn, Benjamin Guggenheim, retrieved 2026] Benjamin Guggenheim LinkedIn post | https://www.linkedin.com/
[LinkedIn, Cristian-George Farauanu, retrieved 2026] Cristian-George Farauanu LinkedIn post | https://www.linkedin.com/
Articles about doubleAI
- doubleAI's Maxwell Rewrites Running Code to Cut Compute Costs — The $1 billion startup, backed by NVIDIA and Dell, plans to charge a share of the savings its AI-powered optimizer generates.