The AI Chart Everyone Is Getting Wrong

June 12, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
June 12, 2026

HEADLINES

SpaceX IPO: The Biggest in History

After months of anticipation, SpaceX is conducting the largest IPO in history — and the retail frenzy has been completely off the charts. Bloomberg reports that retail investors submitted more than $100 billion in orders for a deal that was only selling $75 billion in stock, meaning the retail allocation alone was nearly 7x oversubscribed. Priced at a flat $135 per share by SpaceX, the implied valuation is just shy of $1.8 trillion, debuting as the seventh-largest company in the world ahead of Saudi Aramco, Tesla, and Meta. The flat pricing has drawn concerns about day-one volatility, with Reuters running a rare opinion piece warning that retail buyers may get burned — pointing to SpaceX's 2025 financials showing a $5 billion loss on $18.7 billion in revenue. The Goldman Sachs research team, which simultaneously conducted the IPO and published wildly bullish analysis forecasting $474 billion in SpaceX revenue by 2030, drew its own share of criticism. Meanwhile, the big picture question hanging over all of this is what a SpaceX debut means for the Anthropic and OpenAI IPOs to follow — and economist Peter Atwater probably summed up the investor dilemma best: "SpaceX has created an idiot moment. Buy it and it goes down, you were an idiot. Don't buy it and it goes up, you were an idiot."

Bezos' Prometheus Closes at $41B Valuation

Jeff Bezos' AI startup Prometheus has closed its latest funding round at a $41 billion valuation, raising $12 billion with participation from JP Morgan, Goldman Sachs, BlackRock, and Bezos himself. The company is building what it calls an "artificial general engineer" — an AI system designed to design and manufacture anything, including complex equipment like jet engines — and has already hired 150 people across San Francisco, London, and Zurich. Bezos dismissed fears of an AI jobs apocalypse, arguing AI will instead produce a labor shortage: "Even though you're shrinking the number of people needed by 10x, AI will create 10x more opportunities." Prometheus is also reportedly exploring a $100 billion fund to acquire legacy industrial companies — the logic being, as one observer put it, that the physical economy can't be scraped: you don't find the training data, you acquire the factories that generate it.

Meta Cuts Ties with Manus

Meta has completed an operational split with Manus, complying with orders from Chinese officials who had directed the deal to be unwound. Bloomberg reports that Meta has fully firewalled operations: Manus staff can no longer access Meta's data systems, and Meta staff can no longer use Manus tools for internal work. Manus — one of Meta's flagship acquisitions coming into the year, bought for $2 billion — is now attempting to raise $1 billion to fund a buyback, though it's unclear whether any investors are willing to step up. Beyond Manus itself, the fallout has sent a chilling signal through China's startup ecosystem: the red-chip corporate structure, in which Chinese companies decamped to Singapore to court foreign capital, has effectively been shut down. The FT reports that numerous prominent startups including Stepfun, Moonshot, and Kling are now unwinding their offshore structures and reincorporating in China, with an attorney at Wintell & Co. summing it up bluntly: "Whether to dismantle the red-chip structure is no longer in question. The key is how to complete the restructuring as cheaply and efficiently as possible."

Google Looks to Samsung for TPU Manufacturing

Backlogs at TSMC are pushing Google to look at Samsung for parts of its next-generation AI chips. The Information reports that Google is evaluating Samsung's 2-nanometer process for components of its 10th-generation TPUs, codenamed Icefish — specifically the memory input-output die that connects the processor to memory chips. Google has exclusively used TSMC until now, but the Taiwanese chipmaker's years-long waitlist is forcing customers to get creative. What's emerging is a complex, distributed supply chain: TSMC produces the most advanced processor dies, while Samsung and Intel increasingly handle less cutting-edge components. Earlier this week it was also reported that Google had placed orders with Intel for advanced packaging on their 2028 production run. None of this appears to reflect dissatisfaction with TSMC — the wait times are simply forcing chipmakers to look elsewhere to keep pace with demand.

KKR, NVIDIA, and Kuwait Launch $10B Data Center Company

Private equity continues to pile into data center infrastructure as KKR and NVIDIA announced Helix Digital Infrastructure, a $10 billion construction company backed by KKR, Kuwait's Sovereign Wealth Fund, NVIDIA (through chip and infrastructure deployment), and power company Vistra. Former AWS CEO Adam Selipsky is attached to lead the venture, and his pitch captures the underlying logic: "Data centers, power, and connectivity have all too often been built on separate tracks. That fragmentation has become an industry-wide bottleneck." This is one of several similar full-stack deals in recent months — Broadcom announced a comparable tie-up with Apollo and Blackstone earlier this week — as JLL reports nearly half of data center projects nationally are being delayed. Bringing chipmakers, utilities, and capital together in a single vehicle is increasingly looking like the only way to get these projects across the finish line.

Goldman: The AI Infrastructure Boom Is Being Underestimated

Goldman Sachs strategists led by Ryan Hammond are calling current AI capex forecasts too conservative — and by a wide margin. While the median Wall Street analyst expects the industry to deploy $920 billion in AI data center spending next year, Goldman's team is projecting $1.1 trillion as a baseline and $1.4 trillion in a bull case. Their key assumption: AI demand is still in the opening innings, with token consumption expected to grow 24x through 2030 as agents proliferate. They note that Google and AWS now report a combined backlog of $832 billion, up from $358 billion just six months ago — and that at roughly 1.5% of GDP, this year's AI spending still has room to run before reaching the peaks of past infrastructure waves like railroads and electrification. Goldman's bottom line: peak AI spending is nowhere in sight.

MAIN STORY

The AI Chart Everyone Is Getting Wrong

The speed at which Wall Street went from token-maxxing to token panic is head-spinning — and the chart at the center of it all, the Silicon Data LLM Token Expenditure Index, shockingly doesn't say what everyone on social media is saying it says. This episode is about why that chart has nothing to do with token demand, nothing to do with token volume, and nothing to do with actual token expenditure — and why the real signal it does contain is actually the story we've been tracking here for weeks: the shift from the token subsidy era to the token scarcity era.

IT DOESN’T MEAN WHAT YOU THINK IT MEANS

The Chart and the Frenzy
The Citadel note went viral. The panic was mostly confected.
Citadel Securities published a research note called "Tokenomics" featuring the Silicon Data LLM Token Expenditure Index — a big, scary downward line. The response was immediate: former crypto founder Mo Shaikh's post calling it unexpected got over half a million views; Zero Hedge ran a piece called "Tokenomics Equals Panic"; Real Vision's Andreas Steno Larsen declared the chart "the one everyone should be watching," warning the entire hardware and data center trade could be over for this cycle if it rolls over further. The pattern — someone loudly proclaiming that the agentic frenzy of the past three months has somehow reversed — is one to be allergic to, and it warranted a close look at what the chart is actually measuring.

What the Index Actually Measures
It's a weighted average price per million tokens — nothing more.
Silicon Data themselves took to Twitter to clarify: their index should really have been called the "Token Expenditure Price Index," because it measures the usage-weighted average price the market is paying for a million LLM tokens — not total token demand, not total token volume, not total expenditure. What the downward line is actually saying is that the average price paid per million tokens in mid-June had come down from its peak at the start of June to roughly where it was at the start of May. That's it. There is some interesting signal in there, but it's not the demand collapse signal that Wall Street was reading into it.

The Data Source Problem
The index only draws from third-party token routers — whose whole job is to find cheaper tokens.
Even Silicon Data's own interpretation — that the chart reflects declining "marginal willingness to pay" for frontier models — is probably overstated, because of where the data comes from. The Silicon Data index draws exclusively from third-party token routers: the aggregators and API middlemen whose entire purpose is to route workloads to the cheapest capable model. They have no visibility into direct enterprise relationships with OpenAI or Anthropic, which represent the vast majority of actual token expenditure. What the chart is really capturing is that power users of token routers — mostly developers and highly sophisticated AI buyers — are adding cheaper open-source models to their mix. That's a leading indicator of where advanced AI users are headed, not a measure of the overall market.

The Ramp Data: Most Companies Have Barely Started
The median firm is spending $11.38 per employee on AI monthly.
Ramp tracks AI spend across its customers — who skew more tech-forward than the average business — and the numbers put the token panic in sharp relief. The top 1% of firms, the fully AI-pilled, are spending around $7,500 per employee per month. That's the cohort where you'd expect to see caps and efficiency pushes. But the top 10% are at $610 per month, and the median firm sits at $11.38. Not $1,138 — $11.38. When the median company is spending eleven dollars a month per person on AI, the idea that a shift toward more efficient token use represents some kind of demand collapse is hard to sustain. The growth in total AI consumption as those firms move from eleven dollars to even a modest cap will massively dwarf any revenue lost to efficiency optimization at the top end.

What About Drastic OpenAI Price Cuts?
At 70% margins, there's a lot of room to cut and still be profitable.
Reports suggest OpenAI is considering dramatic price cuts as a preemptive move against Anthropic. But analyst Max Weinbach points out that token margins are high enough that a 60% price cut could still be profitable: "Margin is high now for served tokens. They could cut prices by like 60% and still be profitable, in my opinion." Weinbach's estimates of roughly 70% margins on API pricing are consistent with most independent analysis. Cheaper tokens doesn't necessarily mean less revenue — especially when lower prices could unlock significant new volume from companies currently sitting at eleven dollars per employee per month.

What Citadel Actually Said
More nuanced than the screenshots suggested.
For those who actually read the Citadel note rather than the screenshotted highlights, the argument is considerably more measured. They're describing a bifurcation in AI demand — frontier versus everyday use — not a collapse. They explicitly write that they don't expect the most inference-intensive AI to be abandoned, only that it "is likely to be concentrated among a narrower set of firms with the balance sheets to absorb the compute costs, the research depth to deploy it effectively, and most importantly, the operating domain to scale the rewards from solving genuinely hard problems." That's not a bubble-popping thesis. That's an efficiently allocating market — directing the most expensive AI to the buyers who can use it most effectively.

Thierry Borgeat (X) On the Citadel tokenomics note