Daniel Gross's 2024 AGI predictions scored two years later

Summary

Daniel Gross published a list of AGI-era questions on January 14, 2024, framed as open-ended prompts about what a post-AGI world would look like. Two years on, the scorecard is striking.

Infrastructure

The clearest winner was the infrastructure layer, covering chips, power, and physical buildout, not foundation model labs or application software. Nvidia's gross margins roughly doubled from around 30% to 60%, and its market cap added approximately $3.2 trillion from when Gross published his questions. Microsoft, despite committing $80 billion in AI capex under Satya Nadella, returned just 4% over the same period. The market's implicit verdict was that buying the picks-and-shovels supplier directly beat tying capital to a hyperscaler spending its way toward uncertain returns.

Copper and energy

Copper, which Gross flagged as potentially mispriced, moved from $3.75 a pound in January 2024 to an all-time high of $6.61. A single Nvidia GB200 NVL72 server rack uses over 5,000 copper cables connecting 72 GPUs, with copper wire that would stretch two miles if laid end to end. A 100-megawatt data center, modest by current standards, requires around 3,000 tons of copper. Data centers broadly are expected to consume half a million tons of copper annually within a few years.

Energy was an equally clean trade. Vistra returned 321%, the second-best S&P 500 performance in 2024. Constellation Energy tripled. NRG Energy gained 95%. Nuclear surged on the back of major PPAs: Microsoft signed a $16 billion, 20-year agreement to restart Three Mile Island, Google contracted 500 megawatts from Kairos Power, and Meta committed 6.6 gigawatts across multiple nuclear providers. Coal was a lesser but real story. US coal-fired generation rose 13% nationally and 58% in Oklahoma, driven by data center demand in those regions.

Power transformers emerged as the single most acute bottleneck. Lead times exceeded three years in some cases, with a 30% supply shortfall, and transformer costs surged 150% since 2020.

San Francisco

Gross asked whether San Francisco would become the new Detroit. It didn't. Office vacancy fell from 36.9% to 33.5%. OpenAI now holds a million square feet of offices, Anthropic occupies a 25-story tower, and Sierra signed 300,000 square feet. The Bay Area captured 78% of US AI venture capital in 2025. Overall employment in the city remains below pre-pandemic levels, but the AI ecosystem is the dominant growth driver and housing prices have held firm.

Wealth and wage inequality

The data splits in a counterintuitive direction. An IMF working paper from 2025 argues AI could reduce wage inequality, because high-income roles in law, consulting, and executive functions are more exposed to automation than physical trades. The OECD found assemblers saw wage growth of 11.6% while CEOs saw 2.7%. Wealth inequality is moving the other way. The Magnificent Seven now comprise 32% of S&P 500 market cap and drove 42% of total returns in 2025. AI mega-rounds of $110 billion to OpenAI and $30 billion to Anthropic concentrate private wealth among a narrow group of founders and investors.

India and IT outsourcing

Gross estimated that $250 billion of India's GDP exports were effectively GPT-4 tokens, covering coding and IT services work. That thesis is beginning to show up in headcount data. Major Indian IT firms collectively shed 58,000 employees in 2024 and 2025, a sharp reversal from 2021 through 2023 when they added 360,000. Some of that contraction may reflect post-COVID over-hiring rather than AI displacement alone, and broader Indian employment data has not yet shown a cliff.

Labor and reskilling

Programming jobs are declining while software engineering jobs are growing, with demand for AI engineers up 143%. Entry-level hiring at top tech firms has dropped 25% and internship postings fell 30%. The emerging split is between engineers who can operate at higher abstraction, moving from prompt to deployment or from product to DevOps, and those whose value was in narrower technical execution. Gross's question about whether software engineers become machinists has not answered itself yet, but the pressure on the lower end of the coding stack is real.

Deflation and Jevons

GPT-4-equivalent inference costs fell from $20 per million tokens in late 2022 to $0.04 per million tokens by December 2025, a roughly 500x decline in three years, faster than PC compute costs and faster than bandwidth costs during the dot-com era. AI lab revenues are rising sharply as volume surges, and SaaS vendors are adding AI surcharges of 20 to 37% at renewal. Total spending is increasing even as unit costs collapse, a textbook Jevons Paradox.

Geopolitics and Taiwan

No frontier model has been trained on hardware older than five nanometers, making leading-edge nodes decisive. China's best effort, Huawei's Ascend 910C, runs on SMIC's seven-nanometer-class deep ultraviolet process. It is competitive for inference but requires dramatically more chips and energy for training. Brute-forcing AI with lagging-edge chips appears to hit economic walls before achieving parity.

Gross asked for leading indicators of a Taiwan event. China conducted Joint Sword 2024B exercises in October 2024, surrounding the island with coordinated operations. In December 2025, Justice Mission 2025 deployed over 100 aircraft, with 90 crossing the median line, 13 warships, and 27 rockets fired from Fujian, with 10 landing in Taiwan's contiguous zone between 12 and 24 nautical miles offshore. China also dropped the word peaceful from its official language around reunification in its 2026 to 2035 plan. TSMC's Arizona complex is projected to handle 30% of advanced chip production at scale, a hedge, but a partial one.