Chip-compute costs are rising fast — from $50 billion per gigawatt in 2025 to an estimated $60 billion in 2026 and $73 billion in 2027 — because GPUs, servers, wiring, power, memory, cooling, and ops are the bottlenecks. The argument extends that scarcity into equity upside: NVDA remains cheap on a forward P/E around 20, H100s are still monetizing, Blackwell GPU-hour pricing has doubled in seven months, and SpaceX/Tesla are portrayed as the biggest demand winners with AI compute, satellites, robots, and eventually data centers in space. A second thread says the AI race is being pulled by regulation and geopolitics: 1,130 researchers asked the US to slow frontier AI, the White House wants 30-day pre-release review of closed models from Anthropic, OpenAI, Google, and Meta, and slowing the US could hand China an edge. The closing thesis is that agents, crypto rails, and Starlink’s subscriber growth create a multi-trillion path for SpaceX, while TSLA is a call option on that future and X42-style activity suggests the next phase is already starting.
2025 compute cost is $50 billion per gigawatt, 2026 is estimated at $60 billion, and 2027 at $73 billion, implying 20%-21%+ yearly inflation in the price of building compute.
GPUs, servers, wiring, power sourcing, memory, cooling, and ops/maintenance are the bottlenecks; firms building the physical stack are making a lot of money now.
H100s are still not obsolete and are generating more revenue than ever; Blackwell cluster pricing rose from about $2 per GPU-hour seven months ago to $4 now.
Tesla SpaceX Terra fab is described as the biggest building on Earth at 100 million square feet, approved in Texas, with output of 1 terawatt of compute.
That Terra fab is split 25% for 100 million Optimus robots and 75% for SpaceX/Starship, while Elon Musk said SpaceX will have data centers in space starting next year.
SpaceX AI6 and AI7 chips are not expected until 2027-2028 or later, so SpaceX is said to use NVDA in the meantime, exclusively on Vera Rubin in a customized radiation-proof NVL72.
Anthropic has publicly confirmed an in-house chip design team, aiming for a diversified stack across NVDA, AMD, Google TPUs, and Amazon Trainium because leasing gets expensive fast.
The SpaceX CFO is said to have pointed to $100 billion of ARR before year-end, while the segment projects $100 billion of compute from lease by 2030 and quotes revenue scenarios of $70-$85 billion, $90-$110 billion, and $120-$140 billion over three and a half years.
SpaceX is said to be targeting up to 8 GW of compute by end-2027, with a last-look market value of about $1.3 trillion to $1.5 trillion and a valuation path to $1 trillion-$2 trillion if $100 billion AI compute gets a 10x-20x sales multiple.
SpaceX Q2 revenue is given as $7.8 billion, or $31 billion annualized, with the claim that revenue will more than triple in the next five and a half months.
Starlink active subscribers are said to be above 12 million and adding about 2 million every one or two quarters; v3 satellites are called “alien technology” and expected to accelerate subscriber growth.
Gwyn Shotwell is said to have put a $600 billion target on Verizon, AT&T, and T-Mobile by aiming to take massive share, kill dead zones, and win on disaster coverage, bandwidth, and price; the stocks dropped when she said it.
SpaceX’s S1 path is described as moving the trillion-dollar target from 2031 to 2030 and then possibly 2029, with an implied 170.4% CAGR and a claim that “100 billion by year-end” can compound to $1 trillion in 3.5 years.
SpaceX lockup selling is said to be out of sequence after a 7% unlock, with more pressure expected into October, November, and December, but the price has not fallen because insiders may wait for higher levels.
NVDA is described as having the lowest P/E ever at around 20, with the forward P/E also at its lowest ever, which is used to argue the market is pricing AI like a bubble even though the valuation says otherwise.
GOOGL dropped 5% after leadership changes: Deis left his role to become chief scientist focused on AGI and strategy, and the segment says Google lost its top four AI brains across autonomous experiments, drug development, chip development, and human workflow development.
Meta launched Muse code and a new AI agent for large code bases, positioned as competitive with Claude on features and function while aiming to beat it on price because Claude is described as very expensive.
The White House is said to be asking Anthropic, OpenAI, Google, and Meta to submit frontier models for government review 30 days before release, applying to closed models only and exempting open-weights models.
1,130 AI researchers asked the US government to slow automated AI development over fears of recursive self-improvement outrunning human understanding; the counterclaim is that slowing AI hands China the advantage.
AI agent transactions over the last year are said to have reached X42 cumulative transactions, with Base overtaking Solana as #2, Polygon proof of stake at #3, and Base from Coinbase above 150 million AI-agent transactions.
Within 6 months, AI-agent transactions are projected to reach the billions because agents need crypto rails to pay for things; Houston, Florida, North Carolina, Nevada, Austin, San Antonio, and Dallas are cited as staged for cybercabs and robotaxis with hundreds of vehicles ready, but the stock is said to suffer until launch.