No One Understands How Big This AI War Really Is: Why Compounding Intelligence Rewrites Power

Eight factions are fighting a war that will shape the global economy for the next thirty years. People call it a competition. Too small. The weapon is AI, and unlike land or oil, it compounds.

Capture territory or an oil field and you own a fixed asset. It does not grow more oil. Compute is different. Use it right and you train a better model. Better models pull better users. Better users create better data. That data trains a smarter model. Every loop the leader pulls further ahead, and the gap widens faster each time. That is the whole game.

By 2030, two of these eight win. The other six get acquired, commoditized, reinvented, or go bankrupt. Most still do not know the race is ending.

The real weapon

Skip the GPU press releases that often slip. In 2020, GPT-3 training cost roughly $4–12 million. Three years later GPT-4 landed between $40 and $100 million — about a 10x jump. Training costs have grown about 2.4x per year since 2016. Keep that curve and a single frontier run by 2027 can clear a billion dollars. The club that can stay is already tiny, and shrinking.

GPUs matter. The binding constraint is often HBM — high-bandwidth memory. GPUs are muscle. HBM is the nervous system. Without it, a hundred thousand GPUs are expensive space heaters. Whoever controls HBM controls the training ceiling.

Labs abandoned "Chinchilla optimal" efficiency and now train on roughly 100x more data than the old formula recommends. GPT-4 ate about 13 trillion tokens in 2023 — more than two thousand English Wikipedias. The goal shifted: smart enough that hundreds of millions do not switch, cheap enough to serve them daily. OpenAI alone spends an estimated $8 billion a year on inference. Hitting both targets takes money and silicon almost nobody has.

Wright's Law still applies: when cumulative production doubles, unit cost drops 15–20%. Here the unit is capability per dollar of compute. The faction doubling fastest gets cheaper and smarter at once. The window for new entrants closed in 2024. Matching today's frontier already costs billions. Catching a compounding leader after that is nearly impossible.

OpenAI and Microsoft

ChatGPT hit 100 million users in two months — faster than Google Search or Instagram. As of February 2026: about 900 million weekly active users, 50 million-plus paying subscribers, $2 billion in monthly revenue, 15 billion API tokens per minute. Most downloaded app on Earth after more than 770 million installs in 2025. An ads pilot crossed $100 million annualized in six weeks.

Under the hood it bleeds. A $13.5 billion net loss in the first half of 2025. About $14 billion in losses projected for 2026 against roughly $20 billion in sales. Cumulative losses through 2028 near $44 billion. HSBC sees profitability unlikely before 2030 and a ~$207 billion funding shortfall. They spend about $1.35 per dollar earned. Inference hit $8.4 billion in 2025, headed toward $14 billion in 2026. Talent is leaving toward Anthropic. Mira Murati and Ilya Sutskever left. Stock compensation ran about $6 billion in 2025. Nine hundred million users who cost more to serve than they generate is not a moat — it is a five-front defense while compounding rewards whoever runs.

Microsoft does not need to win the model race. It needs to own the floor. Azure, GitHub Copilot with more than 1.8 million paid developers, Phi as its own model line, and about 27% of OpenAI from the 2019 deal. AI revenue at a $37 billion annualized clip, up 123% year over year. Azure grew 40% in a recent quarter. Lock-in is the motive. The weakness is the same relationship that looks like strength: Microsoft is OpenAI's biggest infrastructure provider and biggest customer. Phi exists because that dependency is a risk.

Meta and China

Llama is not charity. Llama 4 has been downloaded over a billion times. Every time open weights close the gap, OpenAI's pricing power shrinks. Open source does not need to be best. It needs to be good enough and cheap enough that no one holds a monopoly. Meta has 3.56 billion daily active people across its apps. The real threat is an AI layer owned by someone else sitting between Meta and those users. Capex for 2026 jumped to $125–145 billion after $72 billion in 2025 — more than half of every revenue dollar into infrastructure. Muse Spark, Meta's first closed model, means two playbooks at once. The permanent cost: once weights are out, you cannot take them back.

China cannot win the chip race. SMIC is stuck near 7nm; TSMC is at 3nm; without ASML EUV that gap does not close. So the strategy is efficiency and price. DeepSeek R1 hit near-frontier reasoning for a confirmed $294,000 training cost. DeepSeek V4 priced output around $3.48 per million tokens versus roughly $30 from OpenAI and Anthropic — under twelve cents on the dollar. China's five-year plan dropped the word "chip" from AI language and pivoted to economic penetration: deploy cheap models globally, build dependency, convert it into leverage. Price-sensitive markets are the win. Flagship compute is not.

Google and Anthropic

Google invented the transformer in 2017. Custom TPUs are on their seventh generation and beat comparable Nvidia setups on throughput and power. Add 25 years of Search, YouTube, Gmail, and Maps — a dataset no one can buy. AlphaEvolve, a Gemini agent inside Google's stack, designed a TPU circuit and recovered about 0.7% of worldwide compute. Gemini 3.1 Pro led 13 of 16 major benchmarks as of late February 2026. Google Cloud jumped 63% year over year to $20 billion in Q1 2026. A $40 billion commitment into Anthropic puts Google inside the industry's trust layer while it still competes at the frontier. Motive: defend a $333 billion advertising and search machine. Weakness: inventing the transformer and letting OpenAI seize the cultural lead. Own the substrate and you do not need every product cycle.

Anthropic bet that auditability beats raw speed for enterprises and governments. Annualized revenue went from $87 million in January 2024 to $30 billion by April 2026. Over 300,000 business customers. Eight of the Fortune 10. Claude Code crossed $2.5 billion annualized. Amazon committed up to $25 billion; Google $40 billion; SpaceX leased 220,000 Colossus GPUs. Claude is the only frontier model on AWS, Google Cloud, and Azure at once. Then the Pentagon signed AI deals on May 1, 2026 with almost everyone — and skipped Anthropic after it refused autonomous weapons without oversight and mass domestic surveillance. DoD labeled it a supply-chain risk. A judge blocked enforcement. The same guardrail that won enterprise trust blocked the largest procurement budget on Earth. Both things are true. For most people today, Claude still feels like the best model to use.

Musk's stack and the referees

SpaceX acquired xAI in February 2026 at a combined $1.25 trillion. Colossus went up in 122 days. By February 2026 Memphis held about 555,000 Nvidia GPUs — roughly $18 billion in silicon, approaching 2 gigawatts, about four times the next dedicated AI site. Target: one million GPUs. Tesla feeds real-world perception from more than 1.2 million FSD cars every day. Starlink crossed 10 million customers in 160 countries. Motive: own compute, model, fleet, data, and connectivity under one authority. Weaknesses: Grok still learned partly from OpenAI models; Terafab custom chips do not ship before 2027; one person has to keep manufacturing, training, launches, and a fab on track. Most ambitious thesis on the board. Hardest path.

Regulators are the eighth faction. The EU AI Act covers 450 million people with fines up to 7% of global turnover. Full high-risk enforcement hits August 2, 2026. The U.S. saw over a thousand state AI bills in 2025. Safety advocates want Western deployment slowed for ethical reasons. China wants Western deployment slowed for strategic reasons. Same outcome. Regulation moves in months. Compute doubles in weeks.

Two winners

Google wins on substrate: TPUs, proprietary data, Android and Workspace distribution, and internal compounding. When Gemini closes the consumer gap with ChatGPT, the cost advantage underneath starts printing at a scale rivals cannot match.

The Musk stack wins if execution holds: physical deployment at Tesla scale, Starlink, half a million GPUs already live, DoD integration, and the option to rent surplus compute — eventually custom silicon — so he profits even when someone else's model leads a round.

OpenAI most likely becomes infrastructure or gets absorbed. Anthropic becomes the safety layer inside a winner. Microsoft stays enterprise distribution. Meta keeps Llama as a hedge. China builds a parallel industry at prices the West will not match. Regulators become a lobbying battlefield.

This is an infrastructure war the way railroads and spectrum were. Nobody remembers who lost those fights. Everyone lives inside what the winners built. Work that costs $500 an hour today can drop toward five or ten dollars of inference within a decade. Your career edge has to be something a model cannot retrieve and apply on demand. Your kids need skills that compound in an AI-native economy. You have roughly two to three years to read where the chips fall. After that, the shape of the next twenty to thirty years is mostly set.

Start paying attention now.

Check the video here.