The Board Room

Deep dives

1. 01

   Meta Goes Proprietary: The Open-Source AI Safety Net Just Disappeared — And Your 90-Day Window Is Open

   The Break

   For two years, Meta's Llama was the gravitational center of open-source AI. Startups built on it. Enterprises used it to reduce vendor lock-in. The conventional wisdom — that open-source frontier models would always be available — just broke. Meta launched Muse Spark, a closed-weight proprietary model from its new Superintelligence Labs, requiring Facebook or Instagram login. The company simultaneously killed the 2-trillion-parameter Behemoth project and installed Alexandr Wang (acquired via $14.3B Scale AI deal) to lead its AI future.

   Meta's 'hybrid strategy' — open-source small models, proprietary best models — is a polite way of saying 'we'll give you commodity capabilities for free while charging for the ones that matter.'

   What Muse Spark Actually Is

   Independent testing ranks Muse Spark top-5 on the Intelligence Index but behind OpenAI and Anthropic on agentic tasks — the most commercially valuable frontier. Meta went an entire year without releasing a model, and emerged with a product that's competitive on reasoning but trailing where enterprise revenue concentrates. The stock jumped 6.5%, but the strategic significance is in Meta's new posture, not the benchmarks.

   Google's Ecosystem Capture Play

   Google's simultaneous release of Gemma 4 under Apache 2.0 with zero commercial restrictions is transparently an ecosystem capture move — and an effective one. But leaders should be clear-eyed: building on Gemma 4 likely creates soft lock-in to Google Cloud and TPU infrastructure, which is precisely Google's intent. The lesson isn't 'trust Google instead of Meta' — it's that open-source AI strategy now requires multi-vendor optionality by design.

   The Distribution Moat Thesis

   Meta's login requirement isn't a product decision — it's a strategic moat under construction. With behavioral data spanning a decade-plus for 3.5 billion users, feeding this into a 'personal superintelligence' creates a personalization advantage no pure-play AI lab can replicate. The threat isn't that Muse Spark is better today — it's that in 18 months it will know each user so intimately that competing assistants feel generic.

   Market Segmentation Is Hardening

   Segment	Leader	Moat	Risk
   Enterprise	Anthropic	1,000+ $1M+ customers	Pentagon blacklisting
   Consumer + Ads	Meta / Google	3.5B users + ad revenue	Regulation, trust
   Agentic Products	Perplexity (emerging)	$450-500M ARR, 50%/mo growth	Platform competition
   Squeezed Middle	OpenAI	Brand, $730B valuation	No clear segment ownership

   ---

   The era of undifferentiated AI model competition is ending. Distribution and domain moats now determine who wins. Your 90-day evaluation window exists because Meta hasn't yet degraded its Llama open-source tier — but the strategic direction is unmistakable.

   What to do

   1. →

      Audit all production systems, fine-tuned deployments, and pipelines that depend on Llama by end of April. Map switching costs to Gemma 4, Mistral, or proprietary alternatives.

      NowMeta's 'we hope to open-source future versions' is non-committal. The frontier open-source tier will degrade.
   2. →

      Architect a model-agnostic abstraction layer into your AI stack this quarter if you haven't already. With 5+ well-funded frontier competitors, single-vendor dependency is unacceptable risk.

      This sprintModel convergence means switching costs should be near-zero — but only if your architecture supports it.
   3. →

      Evaluate Scale AI dependency for data labeling, annotation, or RLHF pipelines. Meta's 49% ownership creates conflict-of-interest risk for competitors using Scale AI's services.

      This sprintScale AI's largest shareholder now competes with your model provider. Data neutrality is no longer guaranteed.
   4. →

      Map your product's competitive position against Meta's 'personal superintelligence' consumer strategy. Identify which of your AI features survive when Meta bundles equivalent capability into 3.5B daily-active endpoints.

      This quarterMeta's distribution advantage in consumer AI is insurmountable — compete on enterprise, vertical, or privacy differentiation.
2. 02

   The Orchestration Layer Is the New Moat — And Model Providers Know It

   The Convergence Signal

   Two of the most credible voices in tech — Vinod Khosla and Databricks CEO Ali Ghodsi — independently identified the same structural truth this week: AI models are dramatically more capable than current deployments suggest, and the binding constraint is context, not capability. Ghodsi put it bluntly: 'there's still a lot of manual labor happening in every organization' because models lack sufficient organizational context to operate autonomously.

   If you've been allocating capital toward model improvement or selection, you're optimizing at the wrong layer. The defensible, high-margin opportunity sits in the middleware that bridges what models can do and what organizations actually deploy.

   Model Providers Are Paying for Distribution

   The most telling signal: Anthropic committed $200M of its own capital into a $1B venture with Blackstone, General Atlantic, and Hellman & Friedman. OpenAI is pursuing identical PE partnership strategies. When model providers start paying for distribution rather than being paid for their technology, the value stack has inverted. Model capability is becoming table stakes.

   The Agent Infrastructure Collapse

   Anthropic's Claude Managed Agents at $0.08/hr is deliberately designed to collapse the build-vs-buy decision. Rakuten reportedly deployed agents across five departments in approximately a week each. This is the classic infrastructure commoditization pattern: Anthropic is doing to agent infrastructure what AWS did to DevOps. If your engineering teams are building custom agent deployment pipelines, they're building on a layer that will be commoditized within 12-18 months.

   Where Value Actually Accrues

   The venture data confirms this thesis: 53% of VC deal volume in 2025 went to vertical AI. Modus (1 year old) raised $85M for AI audit automation. Patlytics raised $40M for AI patent management. These application-layer companies build moats model providers cannot replicate: proprietary workflow knowledge encoded in evolving harnesses that improve with every execution cycle.

   The Perplexity Proof Point

   Perplexity's 50% monthly revenue jump to $450-500M ARR came not from improving search, but from pivoting to agents. Their 'Computer' product — an agent platform that executes tasks — drove the inflection. Cursor's $2B ARR in coding confirms the same pattern: the market pays for AI that does things, not AI that answers questions.

   ---

   The strategic reframe: your AI investment thesis should shift from 'which model wins' to 'who owns the workflow knowledge and context integration that makes models useful.' That's where the durable margin lives.

   What to do

   1. →

      Double investment in the context and orchestration layer this quarter — specifically the infrastructure connecting AI models to proprietary organizational data and workflows.

      This sprintKhosla's 'massive capability-deployment gap' is your competitive opening. The binding constraint is context, not model quality.
   2. →

      Evaluate Anthropic's Managed Agents against your internal agent infrastructure build within 30 days. Run a 2-week proof-of-concept in a non-critical department.

      This sprintAt $0.08/hr, the cost comparison likely favors buying over building. But lock-in implications must be modeled.
   3. →

      Scout vertical AI acquisition targets in your industry verticals before the remaining independents are absorbed or repriced. Set a decision framework by end of Q2.

      This quarterOne-year-old companies raising $85M means the window for acquiring at reasonable multiples is closing fast.
   4. →

      Shift your AI product roadmap from copilot features to autonomous agent workflows. Benchmark your revenue-per-AI-feature against Perplexity's agent-driven 50% monthly ARR growth.

      This quarterThe market pays dramatically more for task completion than for information retrieval. Copilot positioning is yesterday's market.
3. 03

   AI Is Now Contested Strategic Terrain — DeepSeek on Huawei Silicon, Iran Targeting Data Centers, and the FBI Breach

   The Foundational Assumption Just Failed

   DeepSeek V4 — a 1-trillion-parameter frontier model — was trained entirely on Huawei Ascend 950PR chips without a single NVIDIA component. This is the first frontier-class model trained on indigenous Chinese silicon. The foundational assumption of US tech policy — that chip export controls constrain Chinese AI — has been disproven at production scale. Expect escalatory policy responses that further bifurcate the global tech stack.

   Any competitive analysis or strategic plan that assumes sustained Western compute advantage needs to be stress-tested against a scenario where China achieves chip capability parity within 2-3 years, not 5-7.

   AI Data Centers Are Military Targets

   Iran's IRGC published satellite coordinates of OpenAI's $30B Stargate data center in Abu Dhabi with explicit 'complete annihilation' threats. This is the first nation-state threat against AI infrastructure — and it moves AI compute from a civilian technology concern to a geopolitical one. Any organization planning multi-year infrastructure investments must now incorporate geographic diversification and sovereign risk analysis alongside cost and latency.

   Salt Typhoon Hits FBI

   The FBI declared a 'major incident' from a China-linked breach that compromised systems containing 'returns from legal process and personally identifiable information pertaining to subjects of FBI investigations.' The access vector was a commercial ISP. Salt Typhoon's campaign now spans from 2024 telco compromises to 2026 FBI infrastructure penetration — a strategic intelligence collection operation giving China real-time visibility into who the US government is watching.

   China's Hardware Self-Sufficiency Goes Operational

   Alibaba deployed a 10,000-chip Zhenwu data center with China Telecom, demonstrating Chinese AI hardware self-sufficiency at industrial scale. Combined with DeepSeek V4, Western strategy leaders consistently underweight how fast the global AI ecosystem is bifurcating.

   Implications for Your Organization

   • If you operate across US and Chinese markets: Scenario-plan for complete AI supply chain bifurcation
   • If you have government exposure: Your AI vendor stack is now a geopolitical compliance variable
   • If you depend on lawful intercept or telecom infrastructure: Audit your exposure to Salt Typhoon-style targeting immediately
   • If you're building AI compute: Factor physical security and sovereign risk into site selection — the Stargate precedent is now established

   ---

   The through-line: AI has graduated from a technology competition to a strategic terrain being contested by nation-states. Infrastructure siting, vendor selection, and supply chain decisions now carry geopolitical weight that didn't exist six months ago.

   What to do

   1. →

      Commission a geopolitical risk assessment for your AI infrastructure — map geographic concentration, vendor exposure to political designation, and supply chain dependencies on Chinese vs. Western silicon.

      This sprintThe Stargate threat, DeepSeek V4, and Anthropic's Pentagon blacklisting prove that political risk is now a first-order variable in AI infrastructure decisions.
   2. →

      Audit lawful intercept compliance workflows, CALEA-related data stores, and telecom infrastructure dependencies for Salt Typhoon exposure by end of May.

      This sprintFBI breach via commercial ISP means any company touching telecom infrastructure has unquantified counterintelligence exposure.
   3. →

      Develop dual-scenario strategic plans for your AI investments: one where US-China tech stacks remain interoperable, one where they split completely. Present both to the board by Q3.

      This quarterThe bifurcation is accelerating faster than consensus — Huawei chips training frontier models was supposed to be 5+ years away.
   4. →

      Monitor US policy responses to DeepSeek V4 on Huawei silicon quarterly — new export controls or retaliatory measures will constrain procurement options for global organizations.

      WatchThe failure of current controls virtually guarantees escalatory policy responses.