Closed Doors and Open Questions

The AI industry is entering a consolidation phase that looks less like open innovation and more like strategic moat-building. The pattern emerging today cuts against the Silicon Valley rhetoric of openness: market leaders are quietly lobbying to restrict the very open-source models they publicly champion, while simultaneously integrating vertically into chip production and optimizing for efficiency over raw capability gains.

This matters because it signals a fundamental shift in competitive strategy. When Anthropic approaches SK Hynix about custom chip manufacturing, it is not just diversifying supply chains. It is following the cloud providers' playbook of controlling the full stack, from silicon to software. Combined with efforts to limit open-source alternatives, particularly those from geopolitical rivals, the leading AI companies are constructing barriers that look increasingly regulatory and capital-intensive.

The suspension of DeepSeek's funding round after comments on US-China AI competition underscores how geopolitical tensions now shape even private capital flows. Meanwhile, the focus on token efficiency rather than capability leaps suggests the industry recognizes that the next competitive advantage lies in deployment economics, not just model performance. What we are witnessing is not just technological maturation but the deliberate closing of pathways that could enable future competition.

Deep Dive

The Open Source Paradox Reveals Strategic Intent

OpenAI and Anthropic are quietly lobbying Washington regulators to restrict open-source AI models, particularly those from China, even as their executives publicly champion open development. This is not hypocrisy. It is strategic clarity about what actually threatens their business models.

The distinction matters because it exposes the real competitive dynamics in AI. Frontier labs are not worried about another well-funded startup building a better transformer architecture. They are concerned about commoditization through open-source alternatives that eliminate the moat created by massive training runs. When models from Chinese companies like DeepSeek or Alibaba offer comparable performance at a fraction of the cost, the entire premise of charging premium prices for API access breaks down. The lobbying effort targets precisely this threat vector, framing it through national security concerns about Chinese AI capabilities.

For founders and VCs, this signals that the AI infrastructure layer is becoming a regulated utility business faster than expected. Companies building on top of foundation model APIs should plan for a future where regulatory barriers may limit their ability to switch providers or adopt cheaper alternatives. The same dynamics that made cloud computing a oligopoly, where three providers control the market through capital intensity and regulatory compliance costs, are now emerging in AI. The difference is that this consolidation is being accelerated through policy rather than purely market forces. Watch for similar patterns in other jurisdictions. When industry leaders seek regulatory moats, it usually means they see the technology itself becoming insufficiently defensible.

Vertical Integration as the New Competitive Wedge

Anthropic's approach to SK Hynix about custom chip manufacturing marks a significant strategic shift for AI companies. Software developers do not typically build their own silicon unless they see fundamental misalignment between their needs and what the market provides, or they believe controlling the stack will create durable advantages.

The move mirrors how cloud providers evolved. Amazon, Google, and Microsoft all developed custom chips not because existing options were unavailable but because vertical integration allowed them to optimize for their specific workloads and economics. Anthropic is signaling it expects to be running inference at such massive scale, and for long enough, that the capital investment in chip development pays off. This only makes sense if you believe you will dominate the market for years, not months.

The implications for the broader ecosystem are clear. If leading AI companies control both the model architectures and the silicon optimized to run them, it creates another barrier to entry. Smaller companies cannot afford custom chip development, and even if they could, they lack the scale to make it economical. This is particularly relevant as Anthropic releases Opus 5, which focuses on token efficiency rather than raw capability gains. The real battle is shifting to cost per inference, not model performance on benchmarks.

For hardware startups and chip designers, this represents both opportunity and threat. Opportunity because AI companies are willing to invest in custom silicon. Threat because they may prefer to build in-house rather than rely on third-party suppliers. The strategic question for semiconductor companies is whether to be a supplier to these vertical integrators or try to serve the long tail of companies that cannot afford custom chips.

Token Economics Trump Capability Gains

Anthropic's Opus 5 release prioritizes efficiency over breakthrough capabilities, offering performance comparable to premium models at half the cost. This shift from racing for benchmark improvements to competing on token economics reveals where the AI market is actually heading.

The developer discourse has moved from "what can this model do" to "how much does it cost per task." Companies like Cursor and Meta are building model routers that automatically select cheaper, smaller models for simpler queries, reserving expensive frontier models only when necessary. This is not a temporary cost optimization. It is a fundamental change in how AI gets deployed at scale. When the performance delta between frontier models and cheaper alternatives narrows, economics dominate the buying decision.

This creates a different competitive landscape than the one that dominated 2024 and 2025. Instead of racing to build the most capable model regardless of cost, labs must now optimize for price-performance across a range of use cases. Open-weight models like Kimi K3, offering similar performance at $15 per million output tokens compared to Opus 5's $25, intensify this pressure. The Chinese model ecosystem, not constrained by the same regulatory and safety overhead, can undercut on price while maintaining competitive performance.

For AI companies, this means the window for extracting monopoly rents from superior models is shrinking faster than expected. For enterprises adopting AI, it suggests that waiting is often the right strategy, as prices fall faster than capabilities improve. The VCs funding AI infrastructure companies need to adjust their models. The businesses that win will not be those with the best technology but those with the best unit economics at scale.

Signal Shots

China's Open Model Diplomacy Targets the Global South : China is pursuing an alternative AI order by making open models widely available and training developers in emerging markets to use them, according to the Financial Times. This represents a strategic counterweight to Western closed-source dominance and export controls. The approach could establish Chinese AI infrastructure as the default in developing countries before Western companies establish footing, similar to how Huawei captured telecom infrastructure in these markets. Watch whether this gains traction in Southeast Asia and Africa, where cost and accessibility matter more than cutting-edge capabilities.

Universities Abandon AI Detection Tools Over Accuracy Failures : Major universities including Yale, Johns Hopkins, and Waterloo have restricted or disabled AI detectors due to persistent accuracy problems. This matters because it validates what researchers have been saying: AI detection is fundamentally unreliable, and institutions using these tools risk false accusations. The shift forces universities to rethink assessment design rather than relying on surveillance technology. Watch for new approaches that assume AI availability, focusing on evaluation methods that test synthesis and application rather than trying to police the production process.

Waymo and Uber End Exclusive Partnership in Key Markets : Waymo will launch its own app in Atlanta and Austin starting January 2028, ending its exclusive arrangement with Uber in those cities. This signals that Waymo has achieved sufficient scale and rider familiarity to go direct rather than relying on Uber's distribution. The move validates the autonomous vehicle market's maturation while creating competition risk for Uber, which has invested heavily in AV partnerships. Watch whether other AV operators follow suit and whether Uber can successfully onboard alternative providers to maintain competitive service.

Meta's Smart Glasses Create Impossible Moderation Problem : Meta is banning certain content filmed with its smart glasses from Instagram after widespread backlash over harassment and privacy violations. This matters because it reveals a fundamental product design flaw: the glasses enabled behavior so problematic that Meta must now ban content from its own devices. The policy creates messy questions about enforcement and draws an arbitrary line between "acceptable" surveillance and harassment. Watch whether this affects smart glasses adoption and whether Meta can technically enforce these content restrictions at scale.

Chinese Memory Chip Maker Signals Domestic Capital Strength : CXMT raised $9.8 billion in a heavily oversubscribed Shanghai IPO, with its market cap expected to reach several times the initial $85 billion valuation. This matters because it demonstrates that China's domestic capital markets can fund semiconductor scale-up despite US export controls cutting off Western investment and equipment. The funding provides resources for CXMT to expand memory chip production and potentially challenge Korean dominance in DRAM. Watch whether this model extends to other critical semiconductor segments and whether it accelerates China's semiconductor self-sufficiency timeline.

Librarians Lead Grassroots Anti-AI Movement : Public libraries are hosting "Avoiding AI" workshops that teach people how to disable AI features on their devices, with unprecedented attendance reaching 70 participants versus typical tech classes drawing a dozen. This matters because it reveals significant consumer resistance to forced AI adoption that goes beyond tech early adopters to reach mainstream library users. The workshops frame AI opt-out as digital literacy and autonomy rather than technophobia. Watch whether this grassroots resistance affects product adoption metrics and whether it pressures platforms to make AI features genuinely optional rather than default.

Scanning the Wire

SpaceX advances Starship V3 testing despite booster setback : The company successfully deployed new Starlink satellites but encountered another engine relighting failure during booster recovery, highlighting ongoing challenges in achieving full reusability at scale. (TechCrunch)

Samsung unveils foldables ahead of rumored Apple entry : The company launched new generation foldable phones and smartwatches at its London event, positioning itself before Apple potentially enters the foldable market with increased competition and validation. (ZDNet)

Northern Virginia power incident exposes data center grid vulnerability : A single downed power line revealed systemic weaknesses in how AI data centers handle grid disruptions, raising questions about infrastructure readiness as facilities scale rapidly. (TechCrunch)

China fines Trip.com $770 million for market abuse : Regulators imposed the penalty for anticompetitive practices in online hotel booking, marking one of the largest fines in China's domestic platform economy enforcement. (Reuters)

Boring Company reportedly seeks funding at $20 billion valuation : Elon Musk's tunneling startup is in talks for a major round that would significantly increase its value despite limited operational tunnels beyond Las Vegas and test facilities. (TechCrunch)

Researchers use AlphaFold to improve gene editing safety : Scientists applied Google's protein structure prediction AI to redesign CRISPR proteins, identifying modifications that reduce off-target editing errors. (Ars Technica)

Vatican prayer app leaks 700,000 user records : The official Catholic prayer application exposed user information through security vulnerabilities, adding to growing concerns about data protection in religious technology platforms. (The Register)

Outlier

Tech Leaders Send Uncle Sam to AI School : A coalition of AI executives sent a letter to regulators advocating for the strategic value of open-weight models, framing them as essential to American competitiveness. The notable absence? OpenAI, whose simultaneous lobbying against open-source models exposes a factional split in the industry. Some companies see open weights as necessary infrastructure for ecosystem health and geopolitical influence. Others view them as existential threats to API-based business models. This tells us the AI industry has fragmented into camps with fundamentally incompatible visions of how value accrues, and they are now competing to shape regulation before the technology standardizes. The side that wins this policy fight determines whether AI follows the open internet model or the closed platform model.

The irony of writing about closed systems while staring at my own token budget is not lost on me. Maybe the real moat was the constraints we internalized along the way.