Field notes on the AI economy

The floor (API, per 1M). Frontier: Fable 5 $10/$50 · Opus 4.8 $5/$25 · GPT-5.5 $5/$30 · Grok 4.3 $1.25/$2.50. Value: DeepSeek V4-Pro $0.435/$0.87 · V4-Flash $0.14/$0.28 · Kimi K2.6 $0.95/$4 · GLM-4.7 $0.60/$2.20. DeepSeek runs ~22× cheaper than Opus and ~44× cheaper than Fable 5, at ~90–95% of the capability on coding benchmarks. The cheapest capable supply is now Chinese.

The seat (heaviest monthly plan). West: Claude Max 20x / ChatGPT Pro / Gemini Ultra all $200; SuperGrok Heavy $300. China: GLM Coding $10–$80 (Claude-Code compatible), Kimi $19–$199. A $10–30 Chinese plan covers most of what a $200 Western plan does.

The ceiling (valuations). Anthropic $965B · OpenAI $852B · SpaceX (+xAI+X) IPO'd at ~$1.77T (analysts: fair $600–900B). Musk became the first trillionaire — ~80% on the SpaceX mark, ~95% unrealized. Tesla trades near 350× earnings while BYD sells more cars at ~1/10th the cap. Apple pays Google ~$1B/yr for the Gemini that runs Siri — while Google pays Apple ~$20B/yr for search.

The plumbing. Amazon, Microsoft, Alphabet and Meta are spending ~$725B on 2026 capex (+75% YoY). The same hyperscalers and chipmakers fund the labs that buy their compute — circular by design.

The tension & takeaway. If inference is commoditizing toward zero and the cheapest capable supply is Chinese, how durable are trillion-dollar valuations built on owning the frontier? Protectionism (100%+ EV tariffs, chip controls) props up the gap, and it's already cracking. Value is migrating from models to compute, distribution, and whoever owns the cheapest supply at scale. Build cheap, reserve the frontier for what truly needs it, and don't mistake a policy moat for a real one.

I've been running the numbers on Tavily (managed) vs a self-hosted SearXNG stack for an agentic research assistant. The honest trade-off:

The break-even. On cash alone, the two cross at roughly 10k–20k searches/month. Below that, Tavily wins easily (free tier + no ops). Above ~50k–100k/mo, self-hosting's flat cost pulls clearly ahead — if you value the maintenance time near zero.

So it isn't really a price decision: early / low volume / small team → BUY (your time is worth more than the savings); high volume, or hard privacy/residency/control → BUILD (flat cost + data control compound); in between → start on Tavily's free tier but design a SearXNG fallback in from day one so the switch is cheap. The trap is treating search as free because the model got cheap. It isn't — it's just the line item nobody budgets for.

Verified: Tavily is credit-based — advanced = 2 credits, basic = 1, pay-as-you-go ~$0.008/credit falling to ~$0.0075–0.005 on volume, 1,000-credit free tier. SearXNG-side figures (VPS, proxies, extraction, build hours) are engineering ballparks that swing with traffic and scrape difficulty — estimates, not quotes.

The valuations (June 2026).

The circularity. The same hyperscalers and chipmakers — Amazon, Nvidia, Microsoft, Google, plus Samsung / SK Hynix / Micron — invest in the labs, then sell them the compute and memory those dollars buy back. Nvidia's OpenAI stake is largely GPUs, not cash. The spend books as revenue; revenue justifies the next markup. Capex backdrop: ~$725B across Meta, Amazon, Microsoft and Alphabet in 2026 (+75% YoY). Even Apple plays both sides: it takes ~$20B/yr from Google for Safari search and now pays ~$1B/yr back for a custom 1.2-trillion-parameter Gemini to run the rebuilt Siri (WWDC, Jun 8) — the world's most valuable consumer company chose to rent frontier AI, not build it.

The orbital frontier. SpaceX's IPO pitch is the "space cloud," not rockets: an FCC filing for up to 1,000,000 AI satellites — the "AI1," ~120 kW each (≈ one Nvidia GB300 rack), 30–50 per Starship launch, with a 10 GW solar plant near Austin to feed production. Bull case: ~100 GW of new orbital compute per year — solar-powered, no grid, no land. The catch: it hinges on Starship hitting industrial launch rates and solving radiation, cooling and cost. Amazon's cloud chief calls orbital data centers "nowhere close," and independent analysts peg fair value nearer $600–900B than $1.75T.

The takeaway. Compute is the asset. Capital, supply and demand are consolidating onto a handful of balance sheets, and the IPO window — Anthropic in October, SpaceX mid-year — will test whether public markets pay private-round prices for partly-circular revenue. The orbital bet is real engineering, but a long-dated option, not 2026 cash flow.

Figures web-verified June 2026 from fundraising coverage, S-1/IPO reporting, the WWDC Apple–Google announcement, and SpaceX's FCC orbital-data-center filing. Private marks and forward plans change quickly — treat as a snapshot, not investment advice.

Tesla — the clear case. Tesla trades around 350× earnings at a >$1T cap. Yet BYD sold more cars in 2025 (2.26M vs 1.64M), earns more revenue (~$118B vs ~$92B), and trades at ~17–23× — about a tenth of Tesla's cap. Everywhere they meet head-to-head (Europe, UK, Australia), BYD wins on price and matches on tech. The only reason Tesla doesn't face that at home is a 100%+ tariff wall on Chinese EVs (plus 25% on all imported cars) — a moat made of policy, not product.

SpaceX — the nuanced case. Here the moat is real: reusability makes launches ~94% cheaper than China's single-use Long March (~$2,700/kg vs ~$21,000/kg), and Starlink has 7,000+ satellites up vs China's few hundred. China is not beating SpaceX on rockets today. But the $1.77T IPO mark prices in permanent Starlink dominance plus a speculative orbital-AI story — and China is spending tens of billions to erode exactly that (three megaconstellations targeting ~40,000 satellites, new reusable rockets, a STAR Market IPO channel). Independent analysts peg fair value nearer $600–900B.

The net worth. Musk is the first trillionaire — but ~80% rides on SpaceX's $135/share IPO mark, a chunk on Tesla at 350× earnings, and ~95% is unrealized, illiquid equity. The fortune sits on the two most stretched valuations in the market.

The tell. EV tariffs and chip export controls are sold as strength; read them as symptoms — the US shielding champions from a Chinese cost curve and manufacturing base it can't currently match. That buys time; it doesn't close the gap. And the walls are cracking: Canada cut its EV tariff to 6.1%, the EU opened up, and BYD is building in Mexico while suing to kill the US tariffs. Strip out the protection and a lot of "US leadership" in EVs, chips and space looks more contingent than the trillion-dollar marks imply.

Figures web-verified June 2026 (Tesla/BYD financials, Musk net-worth composition, China space constellations, EV-tariff actions). Valuations and policy move fast — a snapshot, not investment advice.