AI revenue is growing faster than any software business in history, but the quality of that revenue remains contested. Simultaneously, the Iran conflict is reshaping oil markets, and state-level wealth taxes are accelerating capital flight from the West Coast.
1. Iran War: Economic Fallout and Off-Ramp Scenarios
Brent crude spiked from $84 to $119 on day 10 of the conflict, settled near $99 at taping. Iran's new supreme leader Moshtaba announced the Strait of Hormuz stays closed as leverage. Polymarket puts U.S. boots on the ground at 27% by end of March, 57% by year-end.
Goldman Sachs revised PCE inflation forecast from 2.1% to 2.9%, core PCE from 2.2% to 2.4%, and cut GDP by 30 basis points with higher unemployment expected. The S&P has compressed from 24x to 21x forward earnings.
Brad Gerstner (Altimeter Capital) argues the Trump doctrine is pragmatic and limited, not neocon nation-building, which makes the conflict shorter in duration. Chamath Palihapitiya notes Trump's comment that the war would "end very soon" moved oil from $120 to $90 nearly instantly, validating market belief in a short conflict. Chris Wright activated IEA member countries, releasing 172 million barrels coordinated with roughly 400 million total from strategic reserves.
David Sacks outlines escalation risks: Iranian tit-for-tat strikes on Gulf oil infrastructure, targeting of desalination plants serving roughly 100 million people on the Arabian Peninsula, potential exhaustion of Israeli air defenses, and the nuclear escalation tail risk.
Chamath's China thesis: 20% of China's total oil consumption comes from Iran and Venezuela. China kept the U.S. summit on the books precisely because it needs the off-ramp. Xi has every incentive to offer a grand bargain in the three-day summit at month-end. The U.S. produces and consumes 20 million barrels per day domestically; this is a manageable problem for the U.S. and an existential one for China and Asia.
2. AI Revenue Explosion: Real or Experimental?
Anthropic hit a $14 billion annualized run rate in February, growing from $1 billion to $14 billion in 14 months, 12x year-over-year, valued at $380 billion. OpenAI ended 2025 at a $20 billion annualized run rate, up from $2 billion in 24 months, valued at $840 billion. Anthropic reportedly did $6 billion in revenue in February alone, more than the annual revenue of Databricks or Snowflake after 12 years.
Gerstner attributes the inflection to models crossing a threshold where agents compete with labor budgets rather than IT budgets. He cites Palantir, the U.S. military, and NVIDIA as examples of full production deployment.
Chamath counters that there is no documented example of sustained positive margin expansion from AI inside a large enterprise at scale. Amazon issued an edict requiring human review of all agent-generated code after AI-written code caused multiple SEV1 outages in AWS. His framework: revenue from tens of thousands of companies paying $200/month for an AI checkbox is not the same as revenue embedded in critical production workflows the way Databricks or Snowflake are. He is personally consuming triple the tokens every three months with no corresponding revenue increase.
Sacks identifies coding assistance as the primary enterprise breakout use case, noting software engineering has always had a structural supply shortage. He argues the labor displacement framing is misleading because the shortage was so severe that AI augments rather than displaces.
On J-curve economics, Gerstner is building a 1-gigawatt data center in Arizona that escalated from an initial $4-5 billion estimate to $50 billion all-in. Using Sarah Fryer's framework of $10 billion annual revenue per gigawatt, the payback period is five to six years to break even, with profit in years six through eight.
3. AI's PR Problem and the Data Center Cancellation Crisis
AI approval in the U.S. polls only slightly above the Democratic Party. Stanford data shows roughly 80% of Chinese respondents believe AI will be more beneficial than harmful; U.S. respondents are in the low 30s. Chamath traces this to CEOs alternating between doomer messaging for fundraising and selling-tokens messaging for revenue, creating public distrust.
Chamath's data on data center cancellations: roughly 40% of protested data centers get canceled. In 2025, approximately 25 data centers representing 5 gigawatts were canceled. In 2026, roughly 100 are being protested; projected cancellations represent 7 gigawatts. Using Fryer's $10 billion/gigawatt/year figure, 2025 and 2026 together represent $120 billion per year in foregone revenue. Sacks identifies EA-funded doomer think tanks with billions in funding, including Future of Life Institute, as a coordinated driver of local opposition and NIMBY campaigns.
4. Millionaire Tax and Capital Flight
Washington State passed a 9.9% surcharge on income above $1 million, effective 2029, projected to affect 30,000 households and raise $4 billion annually for schools, healthcare, and higher education. Howard Schultz announced his departure from Seattle to Miami on the same day. Jeff Bezos left Washington in November 2023.
Hoover Institution modeled California's billionaire tax across 100,000 Monte Carlo runs: 71% produced negative NPV, expected value is a $25 billion hole. The state overestimated billionaire count, underestimated their tax contribution, and overestimated projected revenue. Departing taxpayers paid $3-5 billion annually. Net result: an estimated $2,500 burden per middle-class household across roughly 10 million households.
Sacks flags the Bernie Sanders/Ro Khanna federal wealth tax proposal at 5% per year as likely to become standard Democratic platform by 2028, with Gavin Newsom leaving rhetorical room to embrace it at the federal level.
Key Takeaways
- The Trump-China summit is the most important variable for the Iran off-ramp. China imports 20% of its total oil from Iran and Venezuela, giving Xi Jinping strong incentive to broker a settlement rather than defend Iran. Markets already signaled this by dropping oil $30 on a single Trump comment about ending the war quickly.
- AI revenue is real and unprecedented in scale, but the distinction between experimental budget-testing and embedded production workflows matters enormously for valuation. Amazon's SEV1 outages from AI-generated code illustrate why enterprise adoption in regulated industries is slower than headline revenue numbers suggest.
- The J-curve for frontier AI is approximately $50 billion of capital per gigawatt of data center capacity, with a five-to-six year payback to break even. Better silicon and open-source models can compress this, but profitability is a decade-scale bet.
- Inconsistent CEO messaging, ranging from existential doomer claims to token-selling utility pitches, has driven U.S. AI approval to historic lows and contributed to the cancellation of an estimated $120 billion per year in data center capacity in 2025 and 2026 combined.
- State-level wealth taxes are not raising net revenue. California's billionaire tax is projected by the Hoover Institution to produce a $25 billion hole once capital flight and departing taxpayers are modeled accurately. Washington State is likely to replicate the same outcome.