DeepSeek Panic, US vs China, OpenAI $40B?, and Doge Delivers with Travis Kalanick and David Sacks

Kalanick positions CloudKitchens as ‘AWS/NVIDIA for food’—owning delivery-only kitchen real estate, proprietary automation like the Bowl Builder, and software that lets brands run virtual restaurants. Restaurants prep ingredients in the morning, then leave; robots handle final assembly, labeling, bagging, and handoff via lockers. This decouples labor-intensive, on-demand cooking from actual order timing, cutting costs and enabling hyper-personalized, nutrition-tracked meals.

Sacks and Chamath argue the $6M figure is just the final training run, not total R&D or hardware, which likely exceeds $1B in GPUs. The real breakthroughs: a new RL algorithm (GRPO) instead of industry-standard PPO and coding directly in NVIDIA PTX (low-level) instead of relying on CUDA. These choices likely emerged from compute and export constraints—and demonstrate how scarcity can force more radical optimization than Western teams awash in capital.

Multiple top AI people Sacks spoke with believe DeepSeek heavily trained on OpenAI outputs. DeepSeek V3 sometimes self-identified as ChatGPT-4, strongly suggesting training on ChatGPT data via web scrape or API abuse. OpenAI has told the FT it found evidence of ‘improper’ use. This raises tough questions: How can Azure or other clouds detect and block abusive distillation at scale? How do you enforce IP in a world where any open-sourced frontier model (like o1) can be cloned, hosted, and undercut by the same partners who profit from you?

The group converges on the view that large general models depreciate fast—Gavin Baker’s ‘fastest-depreciating asset’ line is cited. As performance equalizes and prices plummet, moats move to: (1) ‘shims’ that abstract away any single model and allow hot-swapping; (2) proprietary or structurally advantaged data (e.g., YouTube video, Tesla driving data, domain-specific corpora); (3) application-layer products with deep user integration, workflows, and trust. Chamath expects mixture-of-experts and many small specialized models to further erode the advantage of a single giant model.

Kalanick recounts Uber’s China battles, where Didi copied every product feature in weeks, forcing Uber to build a 400-person China team in SF. Over time, once there’s ‘nothing left to copy,’ Chinese companies start leading in innovation—especially visible in food delivery, logistics, and locker-based last-mile systems. DeepSeek, Meituan, and BYD are presented as examples of this transition from imitators to frontier innovators, especially when pushed by export controls and resource constraints.

Within ~10 days, DOGE is claiming ~$1B/day in taxpayer savings via voluntary federal buyouts (≈8 months severance), strict return-to-office orders (to drive attrition), and aggressive lease cancellations of underused office space. Friedberg connects this to Ray Dalio’s framework: the U.S. must cut the deficit to ~3% of GDP to avoid a debt spiral. Chamath frames DOGE as a three-layer ‘onion’: rightsizing headcount, shedding physical real estate, and—most importantly—digging into IT and vendor spending where opaque contracts likely hide trillions in waste.

Kalanick notes that if all California miles went to EV ride-hail, the state would need roughly double its power-generation capacity; even adding 10–20% is a decade-scale challenge. As robotaxis reduce car ownership and parking needs by ~10x, 20–30% of urban land currently used for parking could become surplus, crashing some commercial real estate values while opening opportunities for housing, local energy production, and robotic fleet depots. The bottlenecks may shift from algorithms and chips to electricity, grid upgrades, and repurposing real-world infrastructure.