AI Compute Polarizes to 1% Amid Layoffs and Hardware Race

China-US AI Gap Stabilizes at 6 Months

Anastasios Angelopoulos from Arena shared leaderboard data showing Chinese models like Deepseek 4 and Kimmy 2.6 closing the gap with US proprietary models, but more slowly than expected. The lag now holds steady at about two quarters (6 months), with US labs pulling ahead in first-derivative improvements. This sustains massive capex on next-gen training, as leading models capture users and developers—but risks complacency if differences become imperceptible, allowing Chinese catch-up to erode moats. "The gap was closing 6 months ago faster than it is now... Chinese models stay roughly two quarters behind."

Photonic Interconnects Unlock Low-Latency Scaling

Nick Harris of Lightmatter emphasized interconnects as the next bottleneck for AI clusters of hundreds of thousands of chips. Their photonic chips use light for data movement, delivering hundreds of terabits/sec per chip—equivalent to 200,000 homes' bandwidth. Hyperscalers like Amazon, Microsoft, Google, and Meta are debt-financing massive builds for capacity and token cost reduction. Optics-enabled systems could hit broad availability in 2028, enabling either bigger models at current latencies or near-zero wait times. Harris noted deployments might split: latency-zero for interactive apps, massive models like Anthropic's Mythos for others. "The thing that defines the performance of these systems is really about how you connect them together."

Space Data Centers Bypass Earth Limits

Philip Johnston of StarCloud detailed orbital megawatt-scale DCs leveraging constant sunlight in dawn-dusk orbits, minimizing batteries (1,000x less than Earth solar) with no clouds, weather, or land costs. StarCloud 2 launches soon with 10kW and Nvidia Blackwell chips; they've ruggedized H100s for space (halved mass, radiation shielding). Collaborating with Nvidia on 'Space Reuben 1'—optimized for mass, heat, radiation—for general use. Jensen Huang showcased StarCloud at GTC. Constellations of 88k satellites enable 20GW, scaling to terawatts, linked by lasers; data up/down via RF/optical or direct-to-device via Starlink relays. Risks like debris are low per Starlink's 30k satellite-years data. "Space is definitely the place to do solar for data centers."

Interaction Models Enable Real-Time Agency

The panel dissected Thinking Machines' dual-model system: a fast 276B-param MoE 'TML Interaction Small' (12B active) for live audio/video/text processing (screen watch, room listen, real-time corrections), backed by a slower thinker. Demos showed pronunciation fixes (acai bowls from Brazil) and fact-checks, keeping humans in loop without interruptions. This unlocks 'unlimited agency' via desktop/camera awareness, 3-pedal desks, and tools like Whisper Flow for voice-to-text. Anastasios highlighted proliferation of Anthropic-like 'Mythos' security models. "Real-time desktop and camera awareness as the real unlock."

Compute Explosion and Elite Polarization

New paradigms demand 100x energy: real-time multimodal agents spike inference needs. Panel agreed on 'polarization of compute'—the 1% affording $10M personal DCs (beyond G650 jets), exhausting for others as companies RLHF-automate jobs. Hyperscalers chase capacity; edge cases like Anthropic's Project Luna (Andon Labs retail store fully AI-run) decouple labor from value. South Korea proposes AI-profit citizen dividends amid unrest.

Layoffs Signal 10x Productivity Gap

AI-first workers outpace non-AI by 10x, fueling layoffs: Cloudflare axed 1,100 (20%) despite record revenue; PayPal, Coinbase, Upwork followed. "Exhausting to... get laid off again." Capitalist incentives falter without smooth transitions; abundance future risks unrest. P(doom) round touched Fermi Paradox implications.

"We're decoupling labor from value creation... building towards a world that I think in many ways we're unprepared for." – Philip Johnston

"You're going to need 100 times the energy and that's exactly the bottleneck we're trying to solve for with these space data centers." – Nick Harris

"The 1% aren't going to be able to just afford a G650. They're going to be able to afford $10 million data center for themselves." – Panel consensus

Key Takeaways

• Chinese AI lags US by ~6 months; sustain capex to maintain lead before 'good enough' erodes it.
• Photonic interconnects and space DCs address energy/interconnect bottlenecks for 2028 low-latency AI.
• Interaction models (fast live + slow thinker) enable creepy-valuable always-on agents via multimodal input.
• Compute polarizes: elites get personal megaclusters; masses face job automation cycles.
• Layoffs hit despite revenues—AI 10x's productivity, decoupling labor/value; prepare for unrest/dividends.
• Ruggedize hardware for space (mass/thermal/radiation); laser links for orbital scaling.
• Proliferating 'Mythos'-like models prioritize security in agent era.
• Transition capitalism to AI abundance or risk systemic failure.