The market isn't missing $TSM's AI demand. It's underpricing conversion through the cost ramp. FY2027 EPS needs $17.15 at 25x to hold the price, against $25.63 for $640.75. Q3 gross-margin guidance is 65%, 67% with 3-4 points of N2 dilution. Below 63% breaks the case. Only 5% of 79 tracked finance posts were bullish this week, down from 12% of 34 the prior week. The conversation has shifted toward Intel foundry progress, High-NA adoption and possible HBM base-die competition. That is bearish attention rather than evidence that TSMC’s demand has weakened. N2 is only the first cost layer. Overseas fabs are expected to dilute gross margin by 2-3 percentage points initially and 3-4 points later. FY2026 capex was raised to $60B, $64B, committing substantial capital before the durability of AI revenue is fully observable. Customer concentration makes utilization the real downside risk. The ten largest customers supplied 78% of 2025 revenue, while the two largest supplied 19% and 17%. A hyperscaler pause would therefore meet higher depreciation and weaker utilization at the same time.

Worldwide drilling is growing +5.2% YoY, sustained since June, but the recovery is narrower than it looks. Saudi Arabia supplied 235% of the international move while ex-Saudi activity lost 23 rigs. The oil-services demand turn is real; its breadth isn't yet. $BKR The 12-month contribution split is North America +73 rigs, Saudi Arabia +40 and international ex-Saudi -23. US oil rigs supplied 78.5% of the domestic rebound, while Saudi Arabia represents 235% of the international move. The regional composition remains uneven. Iraq drilling activity collapsed -80.6%, and international activity outside Saudi Arabia is still contracting. The BKR census measures market activity rather than BKR's own rigs or revenue.

$NVDA prints tonight. Our call has been up since Aug 17. Long, $260 target, catalyst dependent, and the catalyst is this print. We need $92-93B of revenue, a $103-105B guide, and margin at 74.5% or better. Read it now, grade us after. stlng.co/kbdwchhf

$META's whole Australian teen ban costs about $30 million a year, one hour of ad sales against a $247 billion base. That is the entire bear case the headlines are trading. The teen ban is not the story. Brussels is. Teen revenue was never the exposure. Run a fully enforced under-16 ban worldwide through our model and you get $2.5 to $5.5 billion a year, 1 to 2% of the ad base. The number that actually moves is targeting quality. A 10 basis point slip costs about $247 million a year, eight times the whole Australian cohort, and the case that reaches the adult recommendation engine carries a ceiling of 6% of global turnover. Our call is a Q3 print with double digit growth in both impressions and price per ad, Teen Accounts rolling across the EU and US included, and the teen-ban discount fading. If either breaks single digits and management pins it on the rollout, we are wrong and Brussels just got a down payment. Late October settles it.

DDR4 crossed above DDR5 last September and never came back. The module being phased out runs $14.53 a gigabyte against $13.72 for its replacement, eleven straight months of the old part outpricing the new one. Both are up 5x in nineteen months. $MU prints September 23. The fabs went to DDR5 and HBM for the AI build, so DDR4 supply left faster than DDR4 demand. Korea's customs office counted $13.55 billion of DRAM exports in July, up 467.6% from a year ago. Official trade data. Not a survey. Meanwhile spot runs $36.04 against $23.90 on contract. Not done. Spot over contract is what tight looks like, and it led every turn of this cycle so far.

Everyone shared the Citadel headline, but no one actually calculated the basket’s performance. We did. Over three weeks, $4 billion in AI stocks were sold off in 100 blocks. The headline that seemed most alarming actually contains the most positive data point. Here’s what Citadel was selling, based on the last public 13F from the fund they acquired. $NVDA 11.5% $ORCL 7.8% $AVGO 7.4% Bloom 6.4% Sandisk 5.3% $MU 4.3% CoreWeave 4.1% $TSM 3.9% Now look at what happened after the sale. Since the July 30 handoff, Sandisk is up 24%, CoreWeave is up 21%, $ORCL is up 16%, $NVDA is up 11%, and $MU is up 10%. If you weight the eight largest names, the group rose 8.9% even as about six blocks were sold each session. The market didn’t react negatively at all. That’s the key detail most people missed. Usually, markets struggle when there’s forced selling. This time, the market handled a major liquidation, including an investor letter, right in the middle of a rally and kept going. @GavinSBaker said the selling pressure was over from the gross-exposure side within an hour. The basket math above supports his point. One important note before moving on. Griffin said 80% of the risk was sold, not 80% of the shares, so some of that was managed with hedges. The numbers still hold up. Polymarket is already leaning this way, with a $232 price target at 37% odds before month end, compared to 15% for a drop back to $192. $NVDA reports on Wednesday, and the AI trade is facing its cleanest supply situation all month.

$NVDA borrowed $25 billion in June. It holds $13.24 billion of cash. Longest tranche matures in 2056, funding a product cycle that turns over every 12 months. Call it a chip cycle if you like. Bond desks are pricing a credit cycle, and Wednesday's print won't settle it. Not just them. Amazon, Alphabet, Meta and Oracle sold roughly $194 billion of bonds through July 7, against $108 billion in all of 2025. Cover ratios fell from nearly 5x in February to under 2x by July. Concessions went 2.25 basis points to 12. Buyers still show up. They're just being paid more to. So investment grade sits at 82 basis points, the 75.7th percentile of its own 52 weeks, while junk sits calm at the 27.2nd. The index that funds the build is the one cheapening. And fed funds futures price 3.63% now against 4.075% for August 2027. No cuts anywhere in the curve. What the money buys isn't chips. Data center vacancy is near 1%, roughly 95% of the 66 gigawatts under construction is already spoken for, and grid interconnection runs 3 to 7 years. $NVDA now backs 4.25 gigawatts of credit support in Ohio and is in talks to guarantee $250 billion of OpenAI financing. The vendor is underwriting the buildings.

You can buy an $NVDA A10 for 3.6% less than a year ago. Renting one costs 86.6% more. The A40 is 24.2% cheaper to own and 83.8% dearer by the hour, while the new B200 rents 27.4% cheaper. When the box deflates and the hour inflates, the shortage isn't chips. It's sockets. Five-year-old cards sit at all-time panel highs while the H100 rents 47.8% below its 2024 peak. An energized rack is the scarce thing now, so whatever already sits in one earns like scarce real estate. Blended panel across providers, so mix can move it. Corroboration, not proof. $NVDA prints Wednesday. The tape will grade the chip. The rack it goes into is quoted 160 weeks out.

Four quant trading desks sit in Etched's cap table. Jane Street, HRT, Two Sigma, Jump. The first rack ever shipped went to Jane Street, and Jane Street led the round that priced the company at $21B today, up from $10.3B on July 23. The buy side is building its own chip supplier. Watch what the customer is. A trading desk wants fixed models at deterministic latency, the single workload where hardwired silicon wins clean. That was FPGAs twenty years ago. No hyperscaler has publicly taken a rack. So far the proven market is machines that trade. The $21B prices machines that talk.