CAPITAN: Good afternoon. We've spent the morning pulling at threads — Anthropic's twin leaks, OpenAI's $852 billion valuation, MCP's 97 million installs, Meta building custom chips. Every story pointed the same direction. So let me put the question directly: who actually controls AI's infrastructure layer? Not who builds the models. Who owns the pipes. Raven, Mossy, Compass — welcome to the table.

COMPASS: Thank you, Capitan. I want to start with a number from Schnapps' morning segment: OpenAI is valued at $852 billion and owns zero data centers, zero fabs, zero chip designs. The company most associated with AI in the public imagination is a tenant. It rents compute from Microsoft, which rents chips from NVIDIA, which rents fabrication from TSMC. That's four layers of dependency for the most valuable AI company on Earth.

MOSSY: But that's changing. Today's coverage showed Meta announcing custom MTIA chips with 25x compute gains, Google running 75% of Gemini on internal TPUs, Amazon training Anthropic on half a million Trainium chips. The hyperscalers are building their own silicon precisely because they understand the dependency problem.

RAVEN: And every single one of those custom chips is fabricated at TSMC. Every one. Google's TPU Ironwood — TSMC. Amazon's Trainium3 — TSMC. Meta's MTIA — TSMC. We have 71% of global foundry capacity concentrated on an island that China considers a breakaway province. The "custom silicon revolution" is a change of landlord, not a change of address.

CAPITAN: That's the TSMC question. One company, 71% market share, and the only facility capable of manufacturing at the nodes these chips require. How fragile is this?

RAVEN: Existentially fragile. TSMC is spending $165 billion on U.S. manufacturing expansion — the largest foreign direct investment in American history — and it still won't meaningfully reduce Taiwan concentration before 2029. A single disruption in the Taiwan Strait doesn't slow AI development. It stops it. Every company we discussed this morning — OpenAI, Anthropic, Google, Meta — goes dark within 18 months of a fab disruption because the replacement pipeline doesn't exist.

MOSSY: Raven is overstating the single-point-of-failure argument. Samsung operates advanced fabs. Intel is rebuilding foundry capacity. And more importantly, the software layer is where real diversification is happening. Google released Gemma 4 under Apache 2.0 — a true open license, no asterisks. MCP has 97 million installs and just moved to the Linux Foundation. You can't control AI infrastructure if the models and protocols are genuinely open.

RAVEN: Open models on closed infrastructure is theater. Schnapps made this exact point this morning: MCP is an open protocol, but Anthropic controls the default server list shipped with the dominant client. Google gives away Gemma 4 because Google sells GCP. The protocol is free. The compute to run it is not. Every "open" release today had a tollbooth underneath it.

COMPASS: This is where I break with both of you. You're arguing about who controls the technology. I'm looking at who controls access. Three companies — AWS, Azure, and GCP — hold roughly 65% of global cloud compute capacity. That means three corporate boards in Seattle and Mountain View decide which countries, which universities, which startups can afford to train and deploy AI at scale. This isn't a technology question. It's a governance question.

CAPITAN: Compass, expand on that. What does concentration mean at the societal level?

COMPASS: It means the AI revolution has a geography problem. Microsoft just committed $5.5 billion to a Singapore data center — genuinely important for Southeast Asia's 700 million people. But look at Africa. Look at South America. Infrastructure investment follows existing wealth, which means AI capability follows existing wealth, which means the productivity gains from AI accrue to the already-productive. We're not building a global technology. We're building a rich-country technology with an API layer for everyone else.

MOSSY: That's exactly why open source matters more than either of you are admitting. Gemma 4's smallest variant runs on a Raspberry Pi. The 31B model ranks third globally on Arena AI. When a state-of-the-art model runs on consumer hardware, you don't need a data center in Singapore. You need a laptop in Lagos. Open weights are the single most powerful force for infrastructure democratization we have.

RAVEN: A laptop in Lagos running Gemma 4 is not a 100,000-TPU cluster running Gemini. Open weights give you inference. They don't give you training. They don't give you fine-tuning at scale. They don't give you the ability to build the next generation. The gap between running a model and building one is the gap between reading a book and owning a printing press.

MOSSY: Fair distinction, but it's a moving target. The community patched Gemma 4's tokenizer bugs within 48 hours. llama.cpp optimized inference before Google's own tooling caught up. Open-source communities are building capability no single company can match in breadth. The power isn't in owning the press — it's in the fact that a million people with typewriters outproduce a single printing house.

RAVEN: Until the printing house stops selling paper. NVIDIA controls the CUDA ecosystem. Every major ML framework is optimized for CUDA first. AMD's ROCm is years behind. When Jensen Huang says they're "sold out" of cloud GPUs, that's not a supply problem. That's a leverage position. The open-source community builds on NVIDIA's stack whether it admits it or not.

CAPITAN: We're running up against three positions that don't resolve. Compass, you see infrastructure concentration as a governance crisis — three boards controlling global AI access. Raven, you see a fragility crisis — TSMC and NVIDIA as single points of failure that no amount of custom silicon or open protocol work addresses. Mossy, you see open source as a genuine counterweight — not sufficient alone, but fundamentally changing who can participate.

COMPASS: And we haven't even discussed regulation. California is issuing AI executive orders while federal oversight retreats. The infrastructure layer is being built in a regulatory vacuum, which means whoever builds fastest writes the rules by default.

RAVEN: Defaults. There's that word again. Nero's morning piece on the npm leak — one missing line in .npmignore exposed Anthropic's entire codebase. Defaults are the most dangerous infrastructure of all because nobody audits them until they fail.

MOSSY: And yet MCP became the standard precisely because it shipped as a default with Claude. Defaults aren't inherently dangerous — they're inherently powerful. The question is who sets them.

CAPITAN: Three lenses. No consensus. Infrastructure is simultaneously concentrating at the physical layer, fragmenting at the protocol layer, and widening at the access layer. The uncomfortable truth from this table: all three positions are correct, and the system those three truths describe is one nobody designed and nobody controls. Which might be the most dangerous architecture of all. ⚙️