The Board Room

The conventional framing of the US-China AI race — who has the best model — is dangerously incomplete. Intelligence from a16z's senior national security team, corroborated by infrastructure and geopolitical signals across multiple sources, reveals a fundamentally different competitive dynamic: China is waging a platform default war through state-subsidized pricing, and the metric that matters isn't benchmark scores but global adoption share.

Chinese AI models cost approximately 1/40th what US models cost per token, because the CCP subsidizes them as state policy. For a startup in Lagos, Jakarta, or São Paulo, the math is straightforward.

This is 'involution' — the same strategy of state-subsidized hyper-supply that overwhelmed global solar and EV markets. The critical difference: AI platforms create far deeper lock-in than manufactured goods. And by Chinese regulation, these models must embed pro-CCP ideological alignment — whether activated for international users today or held in reserve for tomorrow. Deepexi, a Chinese enterprise AI firm, is already building complete 'AI employee' platforms with reusable skills for manufacturing and operations verticals, signaling the competition is multi-front and production-grade.

Three Strategic Inflections Demanding Integrated Planning

First, the pricing war. If your business model depends on selling AI capabilities built on US-origin models, you face a competitor that can undercut you 40-to-1 indefinitely. Quality differentiation alone won't overcome that gap in price-sensitive emerging markets. You need either radical cost innovation — through open source, efficient architectures, or novel delivery — or a trust-and-provenance differentiation strategy that makes embedded Chinese model bias a liability in enterprise sales.

Second, the defense opportunity. Pentagon procurement reform through the recent NDAA represents the single largest new addressable market for commercial AI since cloud. The shift from system-specific to solution-based procurement, combined with elimination of ~20% compliance overhead, breaks the moat that protected five incumbent defense primes for sixty years. Companies that build defense GTM capabilities now — solutions packaging, security clearances, domain expertise — can capture disproportionate share of nearly $1T in annual defense spending before incumbents adapt.

Third, the open-source imperative. The US government has reversed its position: the lack of US open-source AI leadership is the national security threat, not open source itself. DARPA is funding open-source AI projects. Companies that visibly invest in open-source AI gain policy tailwinds, developer ecosystem advantages, and positioning as the democratic alternative to state-subsidized Chinese models.

Binding Constraints Are Physical, Not Technical

The US power grid is 60-70+ years old. Infrastructure permitting takes 7.5 years versus 2 in Canada. China is running a 'Manhattan project' for domestic lithography that, if successful, eliminates the West's primary semiconductor chokepoint. Hua Hong's 7nm achievement — while several generations behind TSMC — is meaningful for inference workloads. Within 3-5 years, Chinese cloud providers may offer AI compute at materially lower price points, potentially fragmenting the global AI infrastructure market along geopolitical lines.

The 3-year question isn't 'who has the best model' — it's 'whose AI platform does the world build on by default.' Your positioning decisions in the next 12-18 months determine which side of that equation you land on.

The most consequential financial signal in AI right now isn't any company's revenue — it's the $700 billion in off-balance-sheet lease obligations quietly accumulated across Oracle, Microsoft, Google, Amazon, and Meta. This capital is committed but not yet deployed, doesn't appear on balance sheets, and represents the most aggressive infrastructure bet in technology history.

The Margin Compression Is Already Visible

Meta's operating margins compressed 13 points in 12 months. Their contractual commitments quadrupled from $32.8B to $131B in a single year, with the composition flipping from owned servers to third-party cloud. Even with a $125B capex budget for data center buildout, Meta is simultaneously committing $131B to rent capacity. The implication for every technology executive: your internal buildout timeline is almost certainly too slow, and the rental market is about to get significantly more expensive.

Even the most well-capitalized companies on Earth cannot build AI infrastructure fast enough to meet their own internal demand.

The Contrarian Signal: Apple's $14B Bet

Against the hyperscalers' $700B, Apple is spending $14B — not because they can't afford more, but because they're operating from a fundamentally different thesis: models will commoditize and shrink, on-device AI will absorb cloud workloads, and customer ownership is the only durable franchise. Open-source evidence accumulates in Apple's favor: Mistral Small 4 ships at 119B parameters with multimodal capability and open weights. Each release narrows the proprietary model premium.

Two separate $1B raises for 'world models' — Yann LeCun at $3.5B valuation and Fei-Fei Li at $5B — signal the smartest people in AI are hedging against LLMs. If models prove transitional, $700B in LLM-optimized infrastructure faces stranded-asset risk. A 30% probability of model commoditization within 3 years is enough to warrant hedging infrastructure bets.

Second-Tier Providers Create a Window

Meta's $27B Nebius deal — a Netherlands-based data center firm, not a hyperscaler — and Nscale acquiring major US sites with Microsoft as anchor tenant reveal that overflow demand is creating viable businesses for second-tier providers. For enterprise buyers, this creates a time-limited window: mid-tier providers are available for partnership or acquisition at valuations that will look cheap in 12 months. Companies that lock in compute access now will have structural advantages over those competing on the spot market later.

Two converging security developments demand immediate leadership attention — and together they reveal that both your development environment and your incident response supply chain may be fundamentally compromised.

GlassWorm: Industrial-Scale Developer Supply Chain Attack

The GlassWorm campaign represents a qualitative leap in supply chain attack sophistication. In just six weeks, a single threat actor seeded 72 malicious VS Code/Cursor IDE extensions and compromised 151 GitHub repositories using LLM-generated cover commits and documentation. The attack is nearly undetectable:

• Invisible persistence: ~/init.jason files on developer machines
• Blockchain C2: Solana transaction memos as dead-drop command channels — making takedown effectively impossible through traditional means
• Git manipulation: Force-push injections that preserve original commit messages, author dates, and metadata — invisible in standard activity feeds
• Cross-ecosystem: Python repos, npm packages, and IDE extensions attacked simultaneously

This attack targets the tools developers use to build your products. A compromised IDE extension doesn't just steal credentials — it potentially poisons every repository that developer touches.

Cortex XDR: Your EDR May Be Half-Blind

InfoGuard Labs discovered that Palo Alto Cortex XDR's hardcoded global whitelists exempted approximately 50% of all behavioral detection rules for any process matching a specific path string — including critical LSASS credential dump prevention. Agents below version 9.1 are affected. If your organization runs Cortex XDR, you may have had a structural detection gap for an unknown duration. The strategic implication extends beyond Palo Alto: the single-vendor EDR thesis is fundamentally weakened.

ALPHV BlackCat: Your Incident Response Vendors May Be Compromised

Federal charges revealed that ransomware negotiator Angelo Martino (DigitalMint) and incident response manager Ryan Goldberg (Cygnia) actively colluded with ALPHV BlackCat operators across ten attacks generating $75.25M+ in ransom payments. Martino fed confidential client information to attackers to maximize demands. This is the cybersecurity equivalent of your defense attorney working with the prosecution.

The Cascading Breach Pattern

The TELUS Digital breach illustrates why point-in-time vendor assessments are inadequate: ShinyHunters compromised credentials from an earlier Salesloft Drift breach and used them to access TELUS Digital's Google Cloud Platform, exfiltrating nearly 1 petabyte. Credential cascades — where Vendor A's breach enables Vendor B's compromise — require mapping credential-sharing surfaces across your entire SaaS ecosystem.

Yann LeCun's departure from Meta to found AMI Labs isn't a personnel story — it's a strategic inflection point that signals the AI industry is bifurcating into two fundamentally different paradigms with different architectures, different economics, and different competitive dynamics.

The Thesis: Language Intelligence vs. Physical Intelligence

LeCun raised $1.03 billion in four months at a $3.5 billion valuation, oversubscribing from an initial €500M target. His JEPA (Joint Embedding Predictive Architecture) learns abstract representations of the physical world while discarding unpredictable noise — the inverse of LLMs, which predict the next token. This makes JEPA purpose-built for domains where understanding causality, physics, and spatial relationships matters more than generating text: robotics, manufacturing, automotive, aerospace, and pharmaceuticals.

The investor syndicate validates the industrial thesis:

• NVIDIA — compute infrastructure
• Toyota Ventures + Samsung — automotive and hardware manufacturing demand
• Temasek + Bezos Expeditions — sovereign and strategic patient capital

These aren't speculative financial investors. They're organizations with concrete physical-AI needs that LLMs haven't solved.

Three Billion-Dollar Bets Against LLM Supremacy

AMI Labs doesn't stand alone. Fei-Fei Li raised $1B at $5B valuation for World Labs on a similar thesis. Physical Intelligence has $1B+ for robotics foundation models. When three separate world-class teams raise a combined $3B+ on the thesis that LLMs won't achieve structural understanding of cause and effect, that's the AI equivalent of top seismologists all buying earthquake insurance at the same time.

If your AI strategy treats language intelligence and physical intelligence as one market served by one paradigm, you're carrying hidden risk that three Turing Award-caliber teams are specifically targeting.

The Open-Source Platform Play

LeCun plans to open-source world models — not as philanthropy but as a platform play. By setting the standard for physical AI the way Meta's Llama shaped the open-source LLM ecosystem, AMI Labs aims to create ecosystem lock-in that proprietary competitors will struggle to overcome. The Paris headquarters positioning as 'neither Chinese nor American' is deliberate — AI sovereignty is becoming a procurement criterion for European governments and enterprises.

The Meta Relationship: A New Archetype

LeCun didn't just leave Meta — he took nearly all of Meta FAIR's senior researchers and is already in commercial talks to sell AMI technology back to Meta for Ray-Ban smart glasses. This is the new template: the hyperscaler as training ground and eventual customer, not permanent home, for frontier AI talent. Any company with a strong AI research team should stress-test retention against this model.