Advanced Machine Intelligence

1.1 Identity and operating thesis

Advanced Machine Intelligence, or AMI, is a private Paris-based AI research company founded on December 15, 2025. It reports operating locations in Paris, New York, Montreal, and Singapore. AMI describes its mission as building systems that understand the real world rather than relying only on language, using world models that learn abstract representations of noisy sensor data and predict in representation space. The stated application set spans industrial process control, automation, wearable devices, robotics, and healthcare. This is a research-led product thesis rather than a demonstrated commercial model: the assigned evidence describes intended capabilities and application domains but does not disclose a priced product, revenue, ARR, customer count, or current workforce total. The snapshot therefore separates verified identity and financing facts from unsupported operating metrics.[CO001, CO002, CO003, CO004]

1.2 Leadership and execution dependence

Public reporting identifies Yann LeCun as non-executive chairman, Alexandre LeBrun as chief executive, and Laurent Solly as chief operating officer. That configuration combines globally recognized AI research leadership with company-building and operating experience, but it also concentrates execution and external credibility in a small founding group. The near-term plan reported in March 2026 was to add 20 to 30 people, which is a hiring intention rather than evidence of current headcount. Governance diligence is especially important because LeBrun was also reported to remain chairman and chief AI scientist at Nabla while serving as AMI's CEO, and Nabla is AMI's first partner. Investors should verify time allocation, related-party approval, IP ownership, data rights, and commercial terms rather than infer safeguards from the founders' reputations.[CO009, CO010, CO011, CO013, CO014]

1.3 Capitalization and proof burden

AMI reportedly raised $1.03 billion in seed financing at a $3.5 billion pre-money valuation, led by Bezos Expeditions, Cathay Innovation, Greycroft, Hiro Capital, and HV Capital, with strategic participation that included Toyota Ventures, NVIDIA, and Samsung. The round size is financing raised or committed, not evidence of cash currently available, runway, revenue, or product-market fit. The valuation also embeds a substantial execution burden because management's stated horizon for fairly universal intelligent systems is three to five years. Broader market evidence reinforces that distinction: PitchBook and NVCA reported that AI represented 82.7% of U.S. venture deal value through Q3 2026, while Reuters separately reported investor concern that some early-stage AI valuations were frothy and ahead of technical delivery. Those benchmarks are not AMI-specific proof, but they make financing terms, governance rights, milestones, burn, and independent product validation central diligence items.[CO005, CO006, CO007, CO008, CO012, CO015, CO016]

1.4 Exhibits

2.1 Market boundary: world models for consequential physical systems

AMI's relevant market is narrower than all artificial intelligence and broader than humanoid robots. Its investor-authored description centers on world models that represent reality, predict how situations evolve, and connect actions to consequences. Independent reporting identifies prospective uses in autonomous driving, robotics, jet engines, power plants, and models of human organs. That evidence supports a market boundary around predictive and planning systems for physical or safety-critical workflows. It excludes general-purpose text generation, consumer chatbots, and conventional analytics that do not model action in a physical environment. It also separates software intelligence from robot hardware revenue: industrial robots, service robots, autonomous vehicles, and regulated medical devices are adjacent demand pools, not interchangeable spend categories. Status-quo substitutes include scripted automation, domain simulators, human operators, and task-specific control software.[CM001, CM002, CM003, CM011, CM012]

2.2 Sizing: capital intensity is visible, addressable spend is not

The reviewed evidence does not expose a credible AMI-specific revenue pool, contract value, deployment count, or buyer budget from which to calculate TAM, SAM, or SOM. Reuters reported that AI startups raised $73.1 billion in the first quarter of 2025, representing 57.9% of global venture funding, but venture financing is not customer demand and cannot be used as market revenue. The accessible IFR page provides a useful segmentation convention for industrial and service robots, yet its visible underlying data covers 1993-2019 and therefore cannot anchor a 2026 opportunity estimate. The investable conclusion is evidence-constrained: AMI addresses several potentially large adjacencies, while reachable spend remains unquantified until product scope, pricing, procurement pathway, and deployment economics are disclosed. Frothy early-stage AI valuations further weaken any attempt to infer market size from capital raised.[CM004, CM005, CM006, CM012]

2.3 Buyers, users, payers, and the path to budget

Public evidence points to robotics manufacturers, autonomous-vehicle developers, industrial automation firms, aerospace and defense organizations, and scientific simulation teams as plausible buyer groups. The operational user would vary by deployment: robotics engineers, autonomy teams, plant operators, researchers, or clinical-device teams. Budget ownership is not disclosed and likely sits with R&D, engineering, operations, or regulated product organizations rather than a horizontal software budget. AMI said it would focus on research and development during its first year and expected corporate-partner discussions within six to twelve months, indicating a partnership-led discovery phase rather than a mature repeatable sales motion. Named strategic backers may provide domain access, and Nabla is described by one secondary profile as an early deployment partner, but neither signal establishes an arm's-length customer, recurring revenue, willingness to pay, or scalable procurement.[CM003, CM013, CM015, CM016]

2.4 Adoption drivers and constraints

Demand is driven by the gap between fluent generative systems and machines that must reason about consequences, retain context, and act safely over time. AMI's three-to-five year ambition for broadly capable systems creates option value across multiple industries, but the adoption clock should be tied to field reliability rather than the research roadmap. MIT Technology Review reports that humanoid deployment is likely to be slow and industry-specific, with power, battery weight, safety, manufacturing complexity, and absent workplace regulation all limiting scale. Success in one vertical may not transfer to another. Medical applications add a separate regulatory burden: FDA guidance expects planned modifications, validation and implementation methodology, impact assessment, and review through established device pathways. These constraints favor narrow pilots with measurable safety and ROI gates before broad platforms. They also make switching cost and trust endogenous to deployment evidence; neither can be assumed before AMI shows production performance, integration requirements, and buyer economics.[CM007, CM008, CM009, CM010, CM011, CM014]

2.5 Exhibits

3.1 The competitive set extends beyond other world-model startups

AMI is not competing in a single, clean product category. Its stated objective is to build world models that predict how situations evolve and how actions produce consequences, with applications spanning robotics, industrial systems, healthcare, and other sensor-rich environments. That places direct research labs such as World Labs beside robotics-focused Physical Intelligence, incumbent platforms such as Google DeepMind's Gemini Robotics, and vertical applications such as Nabla. It also leaves internal research teams and conventional automation as status-quo substitutes whenever a buyer can solve a bounded workflow without waiting for a general world model. Seedtable's thin profile confirms that AMI is active and positioned as an AI research-and-development platform, but it adds no product-scale evidence. The practical landscape is therefore a contest among generality, deployment maturity, distribution, and trust rather than a list of companies using the same terminology.[CP001, CP002, CP004, CP009, CP011, CP013, CP014]

3.2 Rivals already show different forms of capability and capital intensity

Physical Intelligence is the closest disclosed robotics analogue: CNBC reported a $400 million round at a $2.4 billion post-money valuation in November 2024, while TechCrunch reported in March 2026 that the company was discussing roughly $1 billion more at a valuation above $11 billion after raising just over $1 billion and employing about 80 people. That is capital and talent competition on AMI's own long-horizon terrain, although Physical Intelligence reportedly had no commercialization timeline. Google DeepMind presents a different threat. Its Gemini Robotics paper describes direct robot control, robustness to unseen environments, open-vocabulary instructions, adaptation to novel embodiments, and learning some short-horizon tasks from as few as 100 demonstrations. Those are published capability claims, not a neutral benchmark against AMI, but they demonstrate that an incumbent can combine model research, compute, and an existing platform before AMI has published comparable product evidence.[CP005, CP006, CP007, CP008, CP009, CP010]

3.3 Distribution may matter before model architecture becomes decisive

AMI's first disclosed partnership is Nabla, the healthcare AI company previously led by AMI CEO Alexandre LeBrun. The retained evidence does not disclose the partnership's economics, exclusivity, data rights, or whether Nabla will consume an AMI model in production. Yet Nabla illustrates the distribution advantage of embedding within an established workflow: Hit Consultant reported a rollout to hundreds of University of Toledo Health clinicians through Epic after a pilot, alongside a reported 29% reduction in chart-closure time. Integration into an incumbent system of record can create switching friction and a trusted route to users even when the underlying model layer remains replaceable. Strategic AMI backers including Toyota Ventures, NVIDIA, and Samsung could also become supply, channel, or design-partner advantages, but investment participation alone is not evidence of commercial access, preferred compute, or customer commitments.[CP003, CP011, CP012, CP016]

3.4 The moat is prospective, while commoditization and displacement risks are current

AMI's strongest public differentiation is a coherent thesis: representation-space prediction for continuous, noisy physical data; persistent memory; planning; and an emphasis on controllability and safety. France 24 and RFI reported that the company planned to focus on research and development during its first year, discuss corporate partnerships within six to twelve months, and target broadly useful systems over a three-to-five-year horizon. That timing creates a displacement window. Physical Intelligence can continue scaling robotics models, Google can advance an incumbent robotics stack, World Labs can pursue adjacent real-world foundation models, and vertical applications can own customer workflows before AMI's general platform is commercial. The reviewed corpus contains no disclosed price, contract structure, AMI benchmark, production reliability result, or regulatory validation. The investment case should therefore treat architecture and team quality as options on future differentiation, not proof of present lock-in or pricing power.[CP001, CP004, CP007, CP008, CP009, CP010, CP013, CP015, CP016]

3.5 Exhibits

4.1 Revenue evidence remains pre-commercial

The clearest public commercialization signal is Nabla: multiple reports identify it as AMI's first partner, and a healthcare-industry account describes an exclusive strategic arrangement that gives Nabla first access to AMI's models. That establishes a potential partner-led route into healthcare, but not a revenue stream. None of the reviewed sources discloses a license fee, usage price, minimum commitment, revenue share, contract duration, recognized revenue, ARR, or gross margin. One market profile explicitly characterizes AMI as focused on foundational research with no short-term revenue. The underwriting distinction is important: privileged technology access may validate strategic interest, but it does not establish monetization, cash collection, or revenue quality.[CI007, CI008, CI009, CI010]

4.2 Commercialization is staged behind research

Reported milestones imply a research-first go-to-market motion rather than a scaled sales engine. AMI was expected to focus on R&D in its first year, expand hiring in the near term, and begin corporate-partner discussions within six to twelve months. Its stated horizon for broadly capable systems was three to five years. Those disclosures provide timing proxies, not CAC, payback, sales-cycle, pipeline, utilization, or channel-economics evidence. The likely sequence is capital-funded research followed by selected strategic deployments and only then repeatable commercial packaging. Diligence should test whether Nabla is a paid arm's-length customer, a development partner, or a related-party route to market before assigning revenue value to the relationship.[CI003, CI004, CI005, CI006, CI007, CI008]

4.3 Financing scale does not establish runway

Independent coverage consistently reports a $1.03 billion seed round and an approximately $3.5 billion pre-money valuation. That financing is a major capacity signal, but it is not a disclosed balance-sheet cash figure. The retained evidence provides no closing schedule, tranches, cash-on-hand date, monthly burn, compute commitments, debt facilities, working capital, capex, or planned use-of-funds bridge. Consequently, runway cannot be calculated from public data. Near-term hiring and a multi-year research objective imply material forward spending, yet the magnitude and timing remain private. The correct diligence anchor is a monthly cash bridge from closing proceeds to unrestricted cash, contracted obligations, and milestone-triggered financing needs.[CI001, CI002, CI004, CI011, CI015, CI016]

4.4 Unit economics are unproven and compute-sensitive

No company-specific evidence supports gross margin, contribution margin, deployment cost, inference cost, customer acquisition cost, payback, retention, or service intensity. Broader AI-infrastructure analysis estimates that GPUs represent only half of data-center total cost and argues that commoditized GPU capacity can pressure infrastructure pricing while rapidly depreciating older hardware. Those are context benchmarks, not AMI estimates; AMI may consume compute rather than resell it. The adverse capital-market backdrop also matters: Reuters reported investor warnings that early-stage AI valuations had become frothy and could outrun technical delivery. The financial verdict is therefore research-more: financing reduces immediate fundraising pressure, but revenue quality, margin path, capital intensity, and runway remain blocked by missing company records.[CI012, CI013, CI014, CI016]

4.5 Exhibits

5.1 Product thesis, scope, and maturity

AMI describes its intended product as a new class of AI system that understands the physical world, maintains persistent memory, reasons and plans, and remains controllable and safe. MIT Technology Review provides the clearest independent account of the intended distinction: AMI is pursuing world models for the physical world rather than treating language models as the whole intelligence stack. The disclosed application surface is broad—industrial processes, robotics, autonomous driving, wearables, and healthcare—but it is a portfolio of target workflows, not a set of purchasable modules. AFP reporting carried by France 24 and RFI says management expected the first year to focus on R&D, with corporate-partner discussions on a six-to-twelve-month horizon and more universal systems targeted in three to five years. Those timeframes are management plans, not delivery evidence. The retained public sources do not identify an AMI SKU, API, release, benchmark, production deployment, or customer integration. The underwriting conclusion is therefore research-stage maturity: the technical thesis is specific enough to test, but product-market and engineering execution remain largely private.[CE001, CE002, CE007, CE008, CE015]

5.2 Operating concept, data requirements, and differentiation

The public operating concept begins with video, audio, robot-state, lidar, and other sensor streams, then uses JEPA-style learning to form abstract representations and predict in that representation space while ignoring details that cannot be predicted. AMI has further said that action-conditioned world models should let an agent forecast consequences and plan action sequences subject to safety guardrails. This is a coherent mechanism claim, not a disclosed production architecture: the company has not published model sizes, training corpus rights, evaluation protocols, latency, compute requirements, interfaces, or system boundaries. IEEE Spectrum's account of Covariant illustrates the practical moat AMI would still need to build. Covariant linked its robotics model to tens of millions of trajectories gathered from a deployed warehouse fleet, while acknowledging that the model could not generalize to wholly new objects or situations without suitable data. That comparison is not an AMI benchmark; it shows why proprietary real-world data, deployment feedback, and domain-bounded evaluations may matter as much as the architecture. AMI's public materials do not yet establish any comparable data flywheel or measured generalization advantage.[CE003, CE004, CE005, CE006, CE009, CE012, CE013]

5.3 Deployment, reliability, and trust controls

External evidence argues for narrow, controlled deployment before claims of general physical intelligence. Bain reports that humanoid systems remain mostly in pilots and depend heavily on human input, with structured industrial settings likely to precede open environments; it also identifies battery endurance as a persistent operational constraint. AMI may initially sell software rather than robots, but the same diligence logic applies: a world model must be tested against a defined task, sensor stack, failure envelope, fallback process, and measurable uptime. A Nabla-issued release syndicated through TMCnet says the separate healthcare company completed an Oracle Cerner integration in under 60 days. That is evidence of CEO Alexandre Lebrun's adjacent operating experience, not evidence that AMI has an integration layer or customer deployment. Public AMI evidence states goals of controllability, safety, and guardrails but does not disclose privacy architecture, cybersecurity controls, quality management, incident response, red-team results, regulatory approvals, or certifications. Before underwriting deployment readiness, diligence should require an AMI-specific product demo, architecture and threat-model review, benchmark pack, incident history, release process, and referenceable pilot.[CE010, CE011, CE014, CE016]

5.4 Exhibits

6.1 Disclosed customer proof is limited to one early partner

Nabla is the only named AMI partner in the assigned evidence. TechCrunch, EU-Startups, and Crunchbase News each identify the healthcare AI company as AMI's first partner, while STAT reports that Nabla expects early access to AMI technology. The relationship is substantive enough to create a plausible first validation environment, but it is not yet documented as a production customer deployment: STAT reported no formal equity or licensing agreement between the companies at publication, even though Nabla's COO said they were already working closely. This distinction matters because a partner relationship, early-model access, and shared leadership do not establish a paid contract, production use, renewal, or measurable customer outcome. The reviewed set therefore supports one named partner, not a customer roster.[CU001, CU002, CU003, CU012, CU016]

6.2 Initial segments are hypotheses around real-world operating data

AMI publicly describes target applications spanning healthcare, industrial process control, automation, wearables, robotics, and other systems that operate on continuous sensor data. Those fields imply different buying centers: clinical product leaders and health systems in healthcare, engineering or operations teams in industry, and robotics or autonomy teams for machines in open environments. Only the healthcare path has a named AMI partner. The other segments are application directions, not verified customer categories. A competitor-issued Covariant release shows what mature physical-AI proof can look like - customers across 15 countries and four continents and hundreds of deployed robots - but it cannot be attributed to AMI. AMI's segment map should therefore be read as a prioritized diligence framework rather than evidence of diversified demand.[CU008, CU014, CU015]

6.3 Adoption evidence belongs to Nabla, not yet to AMI

Nabla brings an attractive downstream testing surface: a secondary industry report says its assistant is used by more than 85,000 clinicians across more than 130 U.S. healthcare organizations and names CVS Health, Children's Hospital Los Angeles, and University of Iowa Health Care. The same source reports fivefold Nabla revenue growth over six months. These are Nabla metrics, with Nabla denominators and periods; they must not be relabeled as AMI customer, usage, or revenue metrics. For AMI itself, the assigned sources disclose no active-account count, paid pilot count, utilization, ARR, NRR, GRR, churn, renewal, contract length, or satisfaction measure. Without those measures, durability and expansion cannot be assessed, and the existence of a large partner footprint establishes access to a potential evaluation environment rather than product-market fit.[CU009, CU010, CU011, CU012, CU013]

6.4 Expansion depends on converting research access into arm's-length deployments

Management's own sequencing is long dated. TechCrunch reports that AMI has no plans to generate revenue for the time being, although it intends to engage prospective customers early so models can encounter real data and evaluations. AFP reporting says corporate-partner discussions could begin within six to 12 months, while broadly useful systems are a three-to-five- year objective. That sequence makes early design partnerships plausible but leaves procurement, integration, regulatory acceptance, pricing, and renewal untested. Nabla also concentrates the disclosed customer story in a company whose chairman is AMI's CEO, creating a related-party diligence question without proving any undisclosed transaction. Expansion evidence should require signed arm's-length agreements, AMI-specific deployment outcomes, cohort retention, and customer concentration data before the partnership narrative is treated as a repeatable commercial motion.[CU004, CU005, CU006, CU007, CU015, CU016]

6.5 Exhibits

7.1 Ranked risk posture

The highest residual exposure is research-to-product execution. AMI raised $1.03 billion at a reported $3.5 billion pre-money valuation, yet management said the first year would emphasize research and development, corporate-partner discussions could begin within six to twelve months, and fairly universal systems could take three to five years. That mismatch does not prove failure, but it makes schedule evidence, reproducible safety benchmarks, and partner conversion the decisive monitoring variables. The second tier combines compute procurement volatility, a four-office operating footprint, rapid hiring, and leadership dependencies. Legal and regulatory exposure is potentially severe because factories and hospitals are contemplated applications, while the assigned evidence does not establish approvals, certifications, privacy controls, insurance, or incident-response maturity.[CR001, CR002, CR003, CR004, CR005, CR006, CR007, CR011, CR012, CR013, CR014]

7.2 Regulatory, legal, and safety exposure

Public reporting describes reliability, controllability, and safety as product goals rather than verified operating controls. Nabla is identified as AMI's first partner and a prospective clinical proving ground, which raises privacy, medical-device, model-governance, and professional-liability questions before any production use can be underwritten. Separately, LeCun described AMI's work as a direct continuation of Meta's JEPA research. That lineage is a technical asset, but the assigned sources do not resolve ownership, licensing, employee invention assignments, open-source commitments, or boundaries with Meta and academic affiliations. No absence-of-incident conclusion is warranted: the source pool simply does not contain a regulator record, litigation docket, certification artifact, or incident register.[CR003, CR005, CR008, CR009, CR010]

7.3 Operating, quality, security, and financial-model risk

Compute is a monitorable operating exposure rather than a quantified AMI cost. CloudZero's October 2026 market review put on-demand H100 rental rates at $2.89-$6.98 per GPU-hour with a $3.25 September median, reported a 22% one-month rental increase, and described one-year contract rates as roughly 38% above October 2025. These are market observations, not AMI invoices, so burn and runway remain unknown. Execution complexity also rises with hiring planned across Paris, New York, Montreal, and Singapore. The practical mitigation is staged capacity contracting tied to measured utilization, alongside reliability gates and security-control evidence before safety-critical pilots. Missing supplier commitments, utilization, burn, runway, penetration testing, and recovery metrics prevent a lower residual rating.[CR006, CR007, CR011, CR012, CR013, CR014]

7.4 Partner dependencies, mitigations, and thesis-break triggers

Nabla is the only named operating partner in the retained evidence, so it is both a useful validation venue and a concentration signal until additional counterparties are documented. Leadership overlap compounds that exposure: Alexandre LeBrun moved from Nabla CEO to AMI CEO while remaining Nabla's chairman and chief AI scientist. Financing capacity is substantial, but Reuters-reported investor warnings that early-stage AI valuations were becoming frothy sharpen the downside if technical milestones do not convert into independent deployments. Monitorable triggers are therefore: no documented partner discussions after the stated six-to-twelve-month window; no audited safety or reliability benchmark before regulated pilots; compute commitments without utilization reporting; material schedule slippage against the three-to-five-year ambition; or unresolved IP, data-rights, and governance boundaries. Any two persisting through the next financing decision should pause new capital.[CR001, CR004, CR005, CR008, CR009, CR015]

7.5 Exhibits

8.1 Investment view and recommendation

The recommendation is research-more, not invest-now. AMI has exceptional financing capacity for a seed-stage company: independent reports place the round at $1.03 billion and the pre-money valuation at about $3.5 billion. That scale can fund a serious research program, but it also moves much of the expected technical and commercial success into today's entry price. The retained evidence describes a first year centered on research and development, with corporate-partner discussions expected within six to twelve months, rather than verified revenue, retention, margin, or customer economics. Confidence is medium because the financing facts are consistent across the assigned sources, while the underwriting inputs that determine effective price and return remain private. Risk is high and the valuation stance is expensive until milestones and terms can be verified.[CV001, CV002, CV003, CV004, CV005, CV015, CV016]

8.2 Financing context and comparable signals

The closest public signals describe a market willing to finance frontier physical-AI research before commercialization. Physical Intelligence was reportedly discussing a roughly $1 billion raise above an $11 billion valuation while acknowledging no commercialization timeline and an appetite for additional compute. That is useful as a market-temperature indicator, not as a direct operating comparable: the source does not normalize technology scope, ownership, revenue, burn, or round terms. Broader venture evidence is similarly hot and concentrated. SiliconANGLE's account of PitchBook data says first-quarter 2026 U.S. venture deal value reached $267.2 billion, with five deals representing 73% and AI accounting for 89%. These figures support caution about mark-setting conditions, not a precise AMI multiple.[CV007, CV008, CV009, CV010, CV011, CV012, CV013]

8.3 Bull, base, and bear framing

A defensible scenario framework must remain milestone-based because the assigned evidence does not establish revenue, cash burn, ownership, dilution, liquidation preferences, or an exit multiple. In the bull case, AMI converts its unusually large capital base into externally validated technical milestones and substantive partner commitments on or ahead of the disclosed six-to-twelve-month discussion window. The base case follows the stated research-first plan but leaves valuation unsupported by commercial proof, preserving a research-more stance. The bear case combines delayed commercialization with a correction in early-stage AI valuations, increasing down-round or dilution risk. No numeric probabilities, return targets, or valuation ranges are supportable from the retained public evidence; investment committee modeling should wait for the missing terms and operating data.[CV002, CV004, CV005, CV006, CV008, CV011, CV012, CV015, CV016]

8.4 Entry discipline, diligence asks, and thesis breaks

Entry discipline should require the capitalization table, fully diluted ownership, option-pool treatment, security type, liquidation preference, participation rights, tranches, side letters, pro-rata rights, and any debt or compute commitments. Commercial diligence should obtain signed partner agreements, revenue and pipeline by stage, customer concentration, retention, gross margin, monthly burn, runway, and milestone-linked compute budgets. Technical diligence should define measurable reliability and safety thresholds, using a documented risk process such as the voluntary NIST AI Risk Management Framework rather than treating broad safety language as validation. Thesis-break triggers are a missed partner window, repeated milestone slippage, financing terms that materially subordinate new investors, or failure to produce independently testable performance and safety evidence before the next capital event.[CV005, CV008, CV012, CV014, CV015, CV016]

8.5 Exhibits