SpaceX prices the largest IPO ever, Bezos unveils an "artificial general engineer," and a scary downward chart sends Wall Street from token-maxing to token-panic overnight. NLW spends the back half dismantling that chart — and explains why the signal everyone's screenshotting isn't the signal they think it is.
The scary token chart isn't measuring demand and it's not measuring total expenditure.
Wall Street has decided the Silicon Data "token expenditure index" proves the agentic boom is over. It doesn't. The chart measures the weighted-average price of a million tokens, drawn only from third-party token routers — companies whose entire job is to find cheaper tokens. It says nothing about total demand, volume, or expenditure. The real story is the shift from token subsidy to token scarcity, and when the median firm still spends $11.38 per employee per month on AI, the growth in total tokens consumed will dwarf any drift toward cheaper baskets. That's not a bubble popping — it's a market rationalizing.
By Thursday's close, retail investors had submitted more than $100 billion in orders for SpaceX's $75 billion offering — enough to fill the entire IPO by itself. SpaceX cut the retail allocation from 30% to 20% and priced flat at $135/share, implying a ~$1.8 trillion valuation that would debut it as the seventh-largest company on earth, ahead of Saudi Aramco, Tesla, and Meta.
A rare Reuters opinion piece flagged "a serious risk that investors piling into the world's largest IPO will get burned, especially the retail crowd," pointing to a $5B loss on $18.7B in 2025 revenue. For context, Meta did $200B in revenue last year and even an off-year Tesla managed $95B.
Goldman Sachs simultaneously conducted the SpaceX IPO and published research forecasting $474B in revenue by 2030 with its AI division growing a hundredfold. To skeptics, that's less plausibility than an analyst with a clear incentive forecasting "a bajillion dollars in revenue."
Elon Musk was worth just under $700B last month, with more than 60% tied up in SpaceX. The IPO pricing pushes his net worth to $971B — so any meaningful day-one pop tips him over the trillion-dollar line.
NLW pushes back on treating the SpaceX listing as the market's first price on a frontier AI lab. You can't apply anything about Musk to anyone else — he operates in his own vortex — and the 11th-hour pivot to SpaceX as a neo-cloud reframes the whole thing as an infrastructure build-out story with a heavy dose of Elon halo, not a models story.
— Peter Atwater, economist. Economist Peter Atwater's line cuts through the hyperbole from both the bull and bear camps better than either side's narrative.
Jeff Bezos's AI startup Prometheus closed a $12B round (JPMorgan, Goldman, BlackRock, and Bezos himself) at a $41B valuation. The goal: an "artificial general engineer" that can design and manufacture anything, including jet engines — already staffed with 150 people across SF, London, and Zurich. A reported $100B vehicle would buy up legacy industrial companies, because the physical economy can't be scraped — "you acquire the factories that generate it."
— Jeff Bezos. Bezos dismissed jobs-apocalypse fears as the "opposite of reality," arguing AI produces a labor shortage and predicting two-earner households where one earner drops out "because there's going to be so much productivity."
Meta has completed a Beijing-ordered operational split with Manus, firewalling data systems and tools between the two companies. Beijing ordered the $2B acquisition unwound despite Manus relocating to Singapore first, leaving Manus scrambling to raise $1B to fund a buyback — with attention fading amid open-source harnesses like OpenClaw and core agents like Claude Code and Codex.
The Manus crackdown has chilled the "red chip" structure of decamping to Singapore before raising foreign capital. Step Fun has already reincorporated in China ahead of a Hong Kong IPO, with Moonshot (Kimi) and Kling considering the same — and officials are now seizing passports from key AI researchers and executives, previously beyond the pale.
Google is evaluating Samsung's 2nm process for parts of its 10th-gen TPUs (codename Ice Fish) and already placed Intel orders for advanced packaging on its 2028 run. TSMC still makes the processor itself; a complex supply chain is emerging where less-sensitive components go elsewhere. It's not dissatisfaction with TSMC — just capacity-driven wait times.
Helix Digital Infrastructure, backed by KKR and Kuwait's sovereign wealth fund with Nvidia chips and Vistra power, has $10B in committed capital and is led by ex-AWS CEO Adam Selipsky. It's one of several vertically integrated vehicles (Broadcom-Apollo-Blackstone announced a similar tie-up) aiming to fix the fragmentation between chips, power, and connectivity — even as JLL reports nearly half of US data-center projects are delayed.
Goldman strategists call the median analyst's $920B for 2027 "too conservative," projecting $1.1T baseline and $1.4T bullish. Their key assumption: token consumption rises 24x through 2030 driven by agents, and higher input costs push the nominal dollars of CapEx even higher. NLW says he's firmly in the Goldman camp.
NLW's annual ritual: a new chart inflames investors into an AI counter-narrative frenzy, convinced this time the bubble is bursting. The Silicon Data LLM Token Expenditure Index — shared by Citadel Securities — went viral as proof the agentic boom collapsed. It shows nothing of the kind.
Silicon Data clarified its index "should really have been named the token expenditure price index" — a usage-weighted average price for a million tokens, irrespective of model. The viral downward line just means the average price paid in mid-June fell from an early-June peak back to early-May levels. It has nothing to do with total demand, volume, or expenditure.
The index has no visibility into direct customer relationships with OpenAI or Anthropic, where the vast majority of token spend flows. It draws solely from third-party routers — companies whose entire purpose is to route use cases to cheaper, better-fit models. That structurally exaggerates any shift away from frontier models, so it's a leading indicator of advanced users' orientation, not the average buyer's experience.
The shift from assisted to agentic use cases radically increases AI consumption, and with finite tokens, prices rise as demand outpaces supply. Companies that never had to think about token efficiency — or mixed-basket models that route different use cases to different token tiers — suddenly do. The chart simply reflects that transition in progress.
— Citadel Securities, "Tokenomics" note. Citadel's actual note is far less bombastic than the screenshots suggest, arguing AI demand is bifurcating into frontier versus everyday usage and that the most expensive AI will flow to firms with the balance sheets and operating domains to use it best. NLW: that's not a bubble popping, that's a market rationalizing.
Ramp data shows the top 1% of fully AI-pilled firms spending ~$7,500/employee/month, the top 10% just $610, and the median customer only $11.38 — not $1,138, eleven dollars and thirty-eight cents. The spending caps making headlines come only from the most advanced firms; the vast majority aren't remotely close.
If every firm followed Uber and capped at $1,500/month, the market expansion from a median of $11.38 to $1,500 per employee would massively outweigh any revenue lost to companies getting more efficient. NLW says it's very hard to imagine a short- or medium-term scenario where total AI consumed doesn't dwarf the rebalancing toward cheaper tokens.
— Max Weinbach, analyst. On reports OpenAI may slash token prices to preempt an Anthropic price war, analyst Max Weinbach argues margins on served tokens are high — estimates cluster around 70% on the most inference-intensive API tokens — so a price cut likely means customers can't adopt at volume at current pricing, not collapsing economics.
NLW's tell for empty doom-narrative discourse: phrases like "the whole setup depends on this..." or loud claims that contradict everyone's lived experience. When a viral post insists the agentic frenzy has snapped back to a pre-agentic state, ask what the chart actually says — and who's incentivized to misread it.