Deep Dive
The Compute Crunch Defining the Decade
The core constraint in the AI buildout is raw compute, and the numbers are staggering. A gigawatt of compute cost $50 million in 2025, but by 2026 it's estimated to jump to $60 billion — a 20% year-over-year increase that continues into 2027 at $73 billion. The bottleneck isn't just GPUs and servers; it's power infrastructure, cooling systems, and ongoing operations. This is already minting money for people building data centers, installing cooling systems, or handling the electrical infrastructure. GPU rental demand shows zero signs of softening. Nvidia's H100 chips, which many thought would be obsolete by now, are generating record revenue, and Blackwell clusters have doubled in price from $2 to $4 per GPU hour in just seven months. This trajectory suggests the market is willing to pay exponentially more for access to cutting-edge compute.
SpaceX's Trillion-Dollar AI Business
SpaceX just announced a 100-million-square-foot Terafactory in Texas designed to produce a terawatt of compute. The output splits three ways: 25% feeds into 100 million Optimus robots for Tesla, 75% goes into Starship satellites, and SpaceX will deploy space-based data centers starting next year using Nvidia's customized Vera Rubin architecture (called the NVL 72) hardened for radiation. The company's CFO revealed a target of $100 billion in annual AI compute revenue by year-end, with three scenarios outlined: conservative ($70-85B), base case ($90-110B matching Google and Anthropic demand), and optimistic ($120-140B). To hit a trillion dollars in revenue by 2029 requires 170.4% annual revenue CAGR — equivalent to Anthropic's growth rate, the fastest in corporate history. At a 20x price-to-sales multiple on $100B in revenue alone, SpaceX's AI business justifies a $2 trillion market cap. Gene Munster called it plainly: SpaceX is the world's only sovereign AI company, likely to become the most valuable on Earth.
The Chip Wars Accelerate
Every major AI company is now designing proprietary chips to reduce dependency on Nvidia. Google has TPUs, Amazon built Trainium for training, and Anthropic just publicly confirmed an in-house chip design team for their Claude models. Anthropic currently uses a diversified stack of Nvidia and AMD GPUs, Google TPUs, and Amazon training hardware. The reason is simple: leasing compute is expensive when you're scaling to trillion-token models, so owning the silicon stack matters. This mirrors the vertical integration playbook of Apple and Tesla. Meanwhile, Meta is launching Muse Code and new AI agents trying to compete with Claude on features and price — though Meta lacks the brand trust in AI that Anthropic and OpenAI command. The fundamental insight is that as compute costs rise, controlling silicon becomes a strategic advantage worth billions in savings.
The Geopolitical and Regulatory Squeeze
The White House just asked Anthropic, OpenAI, Google, and Meta to submit closed AI models for government review 30 days before release — an exemption exists for open-source models. David Sacks was instrumental in this policy. Meanwhile, 1,130 AI researchers petitioned the US government to slow automated AI development, citing fears that recursive self-improvement is outpacing human understanding. The tension is real: slow AI down and China wins; regulate too heavily and government bureaucrats (who may not understand the technology) become bottlenecks. Slowing development also invites risk of Chinese AI dominance, which the creator sees as dystopian. The policy environment remains volatile, but the underlying message is clear: AI infrastructure and chips are now matters of national security.
Robotics, Crypto Agents, and the Source Code Transition
Tesla is staging hundreds of robotaxis and cyber cabs across Houston, Austin, San Antonio, Dallas, Florida, North Carolina, and Nevada. The infrastructure is expanding faster than deployment, suggesting a massive rollout is imminent. SpaceX is simultaneously dominating AI compute while Tesla is primed to dominate autonomous vehicles — together, they could represent a multi-trillion-dollar opportunity. On the crypto side, Base blockchain has processed 150 million AI agent transactions in the past year, overtaking Solana as the second-largest player. AI agents need crypto because they cannot access traditional banks, making blockchain infrastructure essential to autonomous systems. Finally, Elon responded to commentary that source code is becoming obsolete like assembly language — AI will soon write code, compile it to binary, and interface directly with legacy systems. This transition is happening now and reshapes career paths in software engineering.