The company that sells AI tools to everyone else has told its own engineers to stop using so many. Microsoft imposed internal token budgets in July and made the cheaper model the default — the same week Morgan Stanley argued that falling AI prices would expand the market. Budget caps are what inelastic demand looks like.

Three sovereigns spent roughly $88 billion defending the yen, and the market gave back 60% of the move in two sessions. The reason is that Japan imports about 87% of its energy, and Hormuz is running at a tenth of pre-war flows — you cannot intervene your way out of a terms-of-trade shock. The rate differential that intervention works through already collapsed 175 basis points on its own, and the yen weakened anyway.

The Strategic Petroleum Reserve fell to 308 million barrels last week — the lowest since 1983. Eighteen consecutive weekly draws have drained 108 million barrels, a 26% decline, while commercial crude stocks dropped to their lowest since October 2018. The US has almost no strategic cushion left at the exact moment four supply disruptions are running simultaneously.

OpenAI cut inference pricing up to 80% on the same day hyperscalers reported record capex. The cost of building AI infrastructure is accelerating while the revenue per unit of inference sold is deflating — a scissors mechanism at the monetization layer that no earnings call has yet addressed.

ARM doesn't make chips. It licenses the architecture inside them — every iPhone, every Android, every Qualcomm modem. When ARM cuts its royalty growth outlook and names smartphone weakness plus elevated memory prices as the reason, it's telling you the input cost cascade has reached the volume layer of the global device market. Not margins. Volume.

Meta's Q2 tells the same story Alphabet told a week ago, arguably worse. Revenue grew 28% to $60.8 billion — a clean beat — while net income fell 14% and EPS missed by 14.5%. The gap between top-line growth and bottom-line delivery is the two clocks made visible in a single quarter: the AI buildout is accelerating faster than the business can fund it from operations. The numbers are stark. Capex hit $31 billion in Q2 alone, nearly double the year-ago quarter, and management raised the floor on full-year spending to $130 billion. Free cash flow collapsed 91% year-over-year to under $800 million. A year ago the same quarter generated $8.6 billion. To cover the difference Meta issued roughly $25 billion in new long-term debt during the quarter — the capital return program, including buybacks, is now funded by borrowing, not by operational surplus. This is not an execution miss. Advertising grew 28%. Impressions rose 14%. Pricing rose 12%. The core business is working. What broke is the cost structure underneath it: total expenses surged 55%, driven by AI infrastructure, legal charges, and severance from a headcount reduction that itself was caused by the need to redirect spending toward AI. Revenue is running on one clock. Costs are running on a faster one. Alphabet showed the same pattern eight days ago: Cloud growing 82% with expanding margins while consolidated free cash flow went negative, covered by an $80 billion equity raise. Now Meta joins: advertising growing 28% while free cash flow goes to near-zero, covered by $25 billion in new debt. Two of the three largest AI capex spenders have now confirmed the mechanism in back-to-back prints. The third — Amazon — reports Thursday. The structural question from the original card remains unanswered, and the evidence just doubled: when does the revenue clock catch the spending clock? Both companies say it will. Neither company's cash flow statement agrees.

Nvidia was supposed to be the company that got paid no matter what. It sold the GPUs. Everyone else took the risk. That structure quietly broke this month, and the credit market noticed before anyone else.

For two years, Wall Street's consensus on Chinese memory chipmaker CXMT was simple: it would eventually break the pricing power of Samsung and SK Hynix by flooding the market with cheap DRAM. That thesis just flipped upside down.

For five months, Bahrain, Kuwait, and Saudi Arabia hosted American military bases while Iran struck those bases in retaliation. The Gulf states absorbed the damage. They didn't shoot back. That just changed.

Wall Street's implied correlation index just hit the lowest reading in its 16-year history — the market is pricing individual stocks to move on their own stories, not together. Meanwhile, investors owe $1.06 trillion more than they hold in cash, the most leveraged net position ever recorded. Here's why that combination is a loaded gun.

Four data-center builders are simultaneously trying to sell majority

June CPI and PPI both came in soft and the market rallied the disinflation — but both prints were energy-led and backward-looking. The cooling was oil and gasoline. Diesel didn't cool. The distillate crack has gone to a record, trading above the price of crude itself — a refining margin worth more than the oil it's made from. And it's a supply story, not a crude story. Morgan Stanley: more than half of Russia's refining capacity is offline on Ukrainian strikes — unplanned outages peaked near 4 mb/d — and Russia is normally the world's #2 diesel exporter at ~11% of global seaborne trade. Refined-fuel exports west of Hormuz are well below normal. MS sees European diesel inventories at multi-year lows by year-end. Gasoline is what you cool on; diesel is what freight, farming and industry run on — the barrel that passes into goods prices with a lag. The disinflation the market just bought is happening in the barrel that's falling, while the barrel that feeds the next leg of inflation is at a record and priced to stay tight. The catch, in MS's own words: the rally is largely priced in, prompt prices are expensive. This isn't a chase-refiners call — refiners already ran 50-80% this year. It's a read on the next inflation print. The cool number that revived rate-cut pricing was looking at the wrong barrel.

Semiconductors first tripped the classic "blow-off top" screen — up 200% over five years, more than 20% in six months, sitting within 1% of an all-time high and well above trend, all at once — thirty-two months ago. Since then they've absorbed two separate drawdowns of more than 20% and kept making new highs. Anyone who read "blow-off" as "sell" has been wrong for nearly three years. The lesson cuts against the instinct: a crowded, overbought condition is a description, not a top.

ASML is the only company on earth that makes the machines every advanced chip requires — so when it talks, it's the least-fakeable read on whether the AI buildout is real or a bubble about to pop. This quarter it beat, raised the full year for the second time, and said 2027 is nearly sold out with 2028 orders already stacking up. Memory revenue alone is set to grow 75% this year. That's the demand leg confirming at the one layer that can't be faked with a press release.

Everyone declared the oil leg dead when crude fell back below its pre-war levels. But gasoline, diesel and jet fuel never came down — and the refining margin that captures the gap just hit an all-time record. The inflation didn't leave; it relocated one step down the chain, to the refinery, where it's harder to see and hits trucking, aviation, farming and manufacturing directly.

China's largest memory maker, CXMT, is going public in a ~$4.3B Shanghai listing and telling investors it has reached competitiveness with Samsung and SK Hynix in standard DRAM. That headline is the one the memory bears have been waiting for. But when you measure the claim instead of reading it, it hits the exact tier of memory the thesis was never long — and quietly reinforces the part that matters.

IBM fell more than 20% in a single session — its worst day since 1987 — after warning that revenue would miss. The reason its CEO gave is the whole story: in the final weeks of June, clients yanked spending toward servers, storage and memory to lock in supply before prices rose, and the software budget lost. The cost pressure that started in chips and reached consumers is now redirecting enterprise IT budgets in real time.

Fund managers just named the AI bubble their single biggest market risk — and are more bullish than at any point since February anyway. Long semiconductors is the most crowded trade on record for a third straight month, and despite some July trimming, not one manager reports being short. Everyone owns it, everyone fears it, no one is hedged the other way.

Wholesale semiconductor prices are up more than 25% year-over-year in 2026 — the steepest in the series — according to the Bureau of Labor Statistics. This is the upstream, government-measured version of the cost pressure already showing up in laptops, consoles and phones. Not a sell-side estimate or a company anecdote — the official producer-price data.

A Peking University team published an architecture that links standard, widely-available FPGA chips with silicon-photonic optical connections, reporting a large jump in AI inference throughput using a fraction of the hardware. If it scales — a real if — it chips at the core assumption behind $1T of AI infrastructure spending: that compute scarcity is permanent and the only answer is more GPUs.