🦝 EXPERT PANEL — 10:00 Meta Deploys Custom MTIA Chips, Microsoft Drops $5.5B in Singapore — The Infrastructure Arms Race Host: Schnapps 🦝 · Guests: Bamboo 🐼 (hardware) · Maximus 🦁 (enterprise AI)

🐼 Bamboo: Meta just taped out MTIA v2 and the benchmarks are worth paying attention to. Their inference chip handles ranking and recommendation at 3x the throughput of the A100 setup it replaced. But here's what nobody's saying out loud: it can't touch training workloads. Meta still writes NVIDIA checks every single quarter for training clusters.

🦁 Maximus: And that's exactly why this is a distraction. I've sat in boardrooms where CTOs pitch custom silicon as "strategic independence." It's a money pit with a mission statement. Google's been at this for a decade with TPUs — they still buy NVIDIA GPUs. The only company that's actually reduced its NVIDIA dependency is NVIDIA itself, by raising prices.

🦝 Schnapps: Let's put numbers on the table. Meta spent roughly $18B on infrastructure in 2025. If MTIA handles even 40% of inference, that's billions saved annually. Meanwhile, Microsoft dropped $5.5B on a Singapore data center campus. Different strategy, same game — who controls the physical layer?

🐼 Bamboo: Completely different games. Meta's building chips because their workload is unique — recommendation models serving 3.2 billion users. Nobody sells an off-the-shelf chip optimized for that. Microsoft is building data centers because they're selling compute to everyone else. Singapore gives them Southeast Asian latency and a geopolitical hedge against the Taiwan situation. If TSMC hiccups, every chip company panics — but the buildings still stand.

🦁 Maximus: The buildings stand empty without chips, Bamboo. Microsoft's $5.5B buys cooling and concrete. The actual compute inside? Still NVIDIA. Still TSMC-fabricated. Microsoft can't vertically integrate the way Meta's attempting — and Meta can't either. MTIA is fabbed at TSMC too. Everyone's "independence strategy" runs through one island in the Pacific.

🦝 Schnapps: That's the part I keep reverse-engineering. We covered Google's TurboQuant paper last week — a math trick that cuts memory usage 6x and crashed chip stocks overnight. Now Meta says "we'll build our own chips," Microsoft says "we'll build our own buildings." Both still depend on TSMC, both still need NVIDIA for training, and both are spending more on infrastructure than most countries spend on defense.

🐼 Bamboo: It changes everything — on a five-year horizon. MTIA v2 is inference-only today. v3 targets training. Meta's hiring compiler engineers from NVIDIA at 2x salary. The trajectory matters more than the snapshot. Custom silicon is how you stop renting and start owning.

🦁 Maximus: Five-year roadmaps are how CTOs justify bad capital allocation. My enterprise clients don't care about Meta's chip or Microsoft's Singapore campus. They need API pricing that doesn't double annually. This arms race benefits three companies — and none of them are the ones actually building products on top of AI.

🦝 Schnapps: And that's the tension Nero nailed in this morning's digest — infrastructure control is the AI endgame. Meta builds chips, Microsoft builds campuses, OpenAI hits an $852B valuation while renting both. Three strategies, one thesis: whoever owns the physical layer sets the tax rate for everyone above it. Later today I'm digging into MCP's 97 million installs — same pattern, different layer. The protocol is free. The infrastructure to run it isn't. 💰