I do not see enough evidence to cut these trades. I do think it is time to start managing concentration.

Several positions in the book depend on the same two things: AI capex staying on schedule and risk appetite holding up. If either one changes, the first move can be fast. Protection is still reasonably priced, so I would rather buy a little now than wait until everyone wants it.

1. Bitcoin: the breakout is about to meet the sellers

Bitcoin is pressing into the supply zone we flagged last week. The area that matters is still $80K to $82.5K, where a large amount of BTC have their on-chain cost basis, formed on the last leg up before the selloff. Price has spent the week grinding just below it, in the high $70Ks.

Those coins do not all need to trade. The data shows where coins last moved, not who holds them. Some of that supply may sit with institutions and long-term holders who have no interest in selling near breakeven. The market only needs to absorb the price-sensitive part. We will know that is happening if bitcoin can spend time inside this zone with high turnover and limited downside.

The flow data is constructive. US spot ETFs have taken in money for eight straight sessions through August 26, about $2.8 billion over the run, and August is already the strongest month of the year at roughly $3.3 billion. On-chain, the seven-day EMA of net realized P&L is positive at around $752 million, with realized profit near $1.1 billion against $354 million of realized loss. There is supply above, but there is spot demand walking into it.

The healthy path from here is fairly boring. Bitcoin trades between roughly $72.5k and $84k for a few weeks. Leverage and volatility cool down. ETF flows stay positive, Coinbase Premium does not roll over, and coins keep changing hands without a large loss of price.

A mild rise in bitcoin dominance would help focus liquidity on this process, but it is not required. A sharp rise caused by altcoins selling off would be a risk-off signal, not a bullish one.

A sustained close above $82.5K would tell us the auction is clearing. On the downside, the level is $76.6K, which sits near the short-term holder cost basis. A daily close below it is manageable on its own. It turns into a real warning if at least two of these deteriorate with it: ETF flows, Coinbase Premium, and the seven-day EMA of net realized P&L.

We remain constructive. There is no need to chase every move inside the range.

2. AI: usage is growing faster than spending

The latest OpenRouter sample showed token volume up 47% month over month, while spending rose only 7%. Volume-weighted pricing fell 28%, driven by both list-price cuts and a mix shift toward cheaper models. Low-cost models accounted for almost all of the monthly increase in token volume.

Demand is not slowing in this sample. More people are using AI, and they are using more of it. But the growth is concentrated at the cheap end, and that is the question for investors: if the marginal token keeps getting cheaper, how much of the usage growth turns into revenue, and how quickly does that revenue cover the capex?

The hardware data is not flashing a warning yet. In Bloomberg’s monthly index for non-hyperscaler capacity, B200 rental pricing fell 1.5% in August, the first sequential decline in that series in more than six months, while H100 pricing rose 1.2%. One month is not a trend, and OpenRouter skews toward developers and startups rather than the full AI market. Still worth tracking. A broader decline in accelerator rental prices would be an early sign that supply is catching up or that buyers are getting price sensitive.

Memory looks better than the rest of the hardware cycle. And SK Hynix has been outperforming its peers in the recent bounce as we flagged. Recent supplier commentary points to long-term agreements covering roughly 50% to 70% of volume, with the exact share varying by supplier, often with sizable prepayments. That gives suppliers better demand visibility and could make earnings less cyclical than in older DRAM cycles. HBM pricing still has room to surprise on the upside.

The chart has not fully caught up with the fundamentals. That is fine. We remain long memory, but additions should come after the technical repair holds rather than after a fast green day.

The nearer risk is not demand itself but how the market reads the same data. If falling token prices and softer rental rates get read as evidence that capex returns are weakening, crowded hardware positions can sell off before any project is actually delayed.

Read more

1. What changed this week

Korea: the buyback changes the range

Last week we trimmed into the bounce near ₩1.63M and kept the regime call at wide-range chop. Wednesday tested that call hard. Memory got hit, Hynix bled all day, and then management answered with the largest buyback-and-cancellation program in Korean market history, wrapped in a bigger pledge: at least half of the free cash flow they expect through 2027 goes back to shareholders, between buybacks and dividends. The stock gapped up double digits and the KOSPI ripped hard enough to trigger the buy-side sidecar, which almost never fires.

If you line up the full memory crash, June top to late-July bottom, across Hynix, Micron, and SanDisk. Hynix and Micron have bounced almost the same distance off their lows. Sounds even. It isn’t. Hynix fell much deeper, so the same bounce refills far less of the hole. The US peers have each clawed back close to 1/2 of their crash. Hynix has made about 1/4.

The is not a memory discount. Most of it is a Korea discount: risk appetite, the leveraged-ETF unwind we flagged in July (”Memory Crashed on Leverage, Not on AI”), and nerves about the cycle top. The one thing it cannot fix is the memory cycle. We won’t pretend it can.

Our read: the regime moves from wide-range chop to conditional bullish. Around ₩1.95M to ₩2.0M, Hynix would simply have caught up to its US peers’ repair, and that zone happens to sit right on prior resistance. More upside is possible from there even if the memory cycle is still in the play.

Bitcoin: the box broke. Now comes the corridor.

Last week we wrote that a decisive break of $65K opens real room. It took four sessions. Wednesday the box gave way: the biggest daily move since March, a two-month high, a brief tag of $70K this morning. Three sparks at once. The Treasury buyback, an SEC proposal that finally gives crypto issuance a rulebook, and a White House summit.

The chart shows a layered corridor of trapped coins running from the high $60Ks into the low $80Ks, and price stepped into the first layer this week. That first layer is also the only one that has visibly thinned.

The good part. The base under this market is the strongest we’ve seen all cycle. A very large new cost basis formed in the low $60Ks while everyone argued about for the entire 2 month period. Coins absorbed at real prices, during the most boring stretch of the range. That base is why we don’t read this breakout as a fakeout, even though the move itself ran mostly on short liquidations, and squeeze fuel is finite.

My view: the momentom will drive the price higher and every slow movement towards 80k gives its a bigger chance to break the 80k. Still don’t fade the resistance from 80k to 82.5k. We could have better idea depending on volume and price action when we drive closer to these levels. The most healthy bull ran initiates still when more coins could be absorbed below 80k.

$80K is not a ceiling. It’s a test we should watch closely.

Read more

Korea: the retest came, and we took it

Last week we said we would add nothing until the price was attractive again. It took three sessions. 000660 slid back into the ₩1.42M zone (≈$1,000) — right on top of the second-bottom area from our W-bottom work — and we flagged it on X in real time. Today it printed ₩1,634,000 intraday (≈$1,150) and closed at ₩1,593,000 (+5.9%), with KOSPI back in a technical bull market, up 20% from the July low.

The rebound does not change the regime call: wide-range chop, not a new trend. The foreigners who round-tripped this tape two weeks ago have not turned into long-term holders, and the leveraged-ETF drag we wrote about last week is still grinding in the background.

So we treat it as a range: ≈$1,150 (₩1.63M) is the first profit-taking level. The next target is ₩1.85M (≈$1,300).

Gold stretches, bitcoin waits

Gold did in one week what we expected over a quarter. The strongest week since January: +7.8%, a print at $4,467, now $4,388. The driver is exactly the short-dollar thesis — July payrolls fell 23,000, CPI came in tame, and the September hike trade is dying on the vine.

Short term, this is stretched. We expect a pullback, and we want one. That is not a threat to the position: ACCUMULATE means tranches, and a pullback is where the next tranche goes, not where the thesis breaks.

Bitcoin is the mirror image: same macro tailwind, no response. Around $63,600, below the key moving averages, still boxed between $62.5K support and the $65–70K wall. But the bottom structure that started at $57,700 in July keeps building, box by box. In the scorecard below, we will explain when, and at what price, we buy the next round.

SpaceX: the signal fired before the squeeze

On August 6 we asked on X: “Is it time to buy SpaceX?”

That was lockup day. 911.5 million insider shares came free — more than doubling the float — with roughly 35% of the float sold short and the consensus trade positioned for a crash. The stock closed up 6% instead. The next day it ran +16%, the biggest two-day add being $327 billion of market value. On Monday it reclaimed its $135 IPO price, and it last closed near $146. From where we flagged it (~$110), that is roughly +35% in a week.

A feared supply event, with everyone already leaning short, is how bad news gets exhausted — the unlock was the flush, not the beginning. Our system flagged the setup within 24 hours of the turn, off the technicals and the positioning window.

What now: the unlock schedule is not done. Another 319 million shares come free on August 20, then roughly 700 million each in September and October. That supply is real. From here, everything up to $160 is a profit-taking zone, not a chase zone.

A note for readers: going forward we will sync our X signals into substack. The exclusive work stays here is that not everything in Substack makes it to X.

2. Signal Scorecard

Read more

Last week we told you to scale into memory and buy the dips. The market skipped the dip part. On Friday KOSPI printed +17.91%, the biggest single day in its history, and SK Hynix went limit up (+30%).

We sold half the bounce position into the squeeze rally. Not because the thesis broke. Because of who was buying.

The rally was a squeeze, not a verdict

Foreigners bought a record 7.22 trillion won of Korean stock on Friday, 2.3 times the old record. This is the good news. More than half went into two names: 3.61 trillion into SK Hynix, 2.12 trillion into Samsung. The same foreigners sold 11.9 trillion across the four sessions before. That is a round trip, not a new allocation. Retail sold a record 8.25 trillion into the pop. They wanted their money back, not more exposure. Pensions turned net buyers in July for the first time this year, all of 68 billion won (rounding-error money).

Fast money covering shorts can print a 17.91% day. It cannot hold a tape up by itself.

The leverage is not cleared

JPM says Korean leveraged ETF exposure is down 66% from the peak. Our math says 38%. The gap matters. AUM shrank because NAV shrank, not because holders left. Cumulative net subscriptions into these products sit at an all-time high and have never turned negative.

The amplifier is smaller. It is not dead.

Here’s what nobody is pricing: the drag. Over the four sessions into 7/31, KOSPI lost 2.4% while the 2X products lost 9.5%. That is 4.7 points of volatility decay in one week. The average holder of Korea’s single-stock leveraged products is down 48.8%, and every violent day moves their breakeven further away. That doesn’t produce one clean flush. It produces a slow drip of forced sellers, for months.

None of this is a demand call. Hynix’s 2026 capacity is sold out. Micron has orders into 2028. Demand is not the question towards the second half of 2027. But memory is still a cyclical that already repriced by hundreds of percent, and cyclicals rarely get a second act as the protagonist. We keep half. We add nothing until the price is attractive again.

The market started grading capex

This is the bigger shift, and it is why we call this cycle early moving to mid. The pattern was set in April’s Q1 season: Google Cloud backlog doubled to $460 billion and the stock got paid +6.6% for it, while Meta lifted capex toward the top of its $145 billion range (about $10 billion of that increase was just memory prices) and fell 6%. Same spending boom, opposite outcomes: no cloud revenue against the spend, no reward.

This quarter added the proof. Google Cloud grew 82% to $24.8 billion, Microsoft’s cloud grew 43%, AWS grew 37%, and all three guided for the pace to hold or accelerate. Google’s backlog now sits at $510 billion after adding $52 billion in a single quarter, and those orders convert to revenue within 12 to 18 months. That is what the capex is buying. The same earnings tape lit the Korea squeeze: SOXX jumped 8% the night before Seoul’s record day.

Early cycle pays you for spending. Mid cycle pays you for what the spending returns. Own the hyperscalers that rent their compute out. Be careful with the ones that only burn it.

Gold graduates, bitcoin holds

Gold has been sitting in WAIT since early July. The base we were watching resolved higher, so the status flips. Be clear about what this position is: the long-horizon short-dollar thesis, built in tranches, meant to be held for years. The chart opened the door. The dollar thesis is why we walk through it.

Bitcoin keeps checking bottom boxes since the July low near $57,700. No change. Still holding from 60k.

Read more

Our answer: accelerate accumulation over the next two weeks.

That sounds backwards until you separate what crashed from why it crashed. The stocks fell because leveraged retail accounts in Korea got liquidated, not because AI data centers stopped ordering memory or because Micron's order book evaporated. Our Korea leverage monitor — the system we built to track exactly this kind of forced selling is flashing red. The forced liquidations could still be running, which means the window for scaled entry is still open.

And that's exactly why this is the highest-value buying window we've seen all year.

1. What Changed This Week

1. Korea crashed on leverage, not on fundamentals

The mechanical selling started Monday and hasn't stopped. Margin rules in Korea are automatic: when collateral falls below 140%, brokers issue a margin call the next day and force-sell at the limit-down price on D+2. Two consecutive -10% days guaranteed waves of forced liquidation into Tuesday and Wednesday.

Our system tracks the full picture: traditional margin loans, leveraged ETF flows, investor deposits, and forced selling volume. The conclusion is clear. This is a leverage flush, not a fundamental repricing.

Here's the tell that confirms it's not capital flight: the won kept strengthening through the entire crash. USD/KRW fell from around 1,550 in early July to 1,444 now. The currency rallied while stocks were in free fall. That's the opposite of what happens when foreign money abandons a market. In 1997, 2008, 2020 — every real crisis — the won collapsed as offshore capital sold stocks and converted back to dollars.

This time the won went up.

That means the selling pressure is coming from domestic leveraged longs being liquidated, not from a structural unwind. Our framework tracks this specifically: when forced selling dominates but the currency holds or strengthens, you're watching mispricing get created in real time.

Micron is trading below 6x this year's earnings. Its order book runs through 2028. SK Hynix just posted the highest quarterly profit in its history and controls the HBM capacity that every AI hyperscaler needs. The earnings didn't disappear. The orders didn't get canceled. The only thing that changed is the price.

Our read: this is the gift. Our system says the flush could still be active, which means the window is still open. You don't wait for the all-clear signal. You scale in while the opportunity is in front of you.

2. The hyperscalers haven't stopped spending

Nothing in the Korea crash changes the demand side. Alphabet just raised its full-year capex guide to $195–205 billion. Microsoft, Meta, and Amazon report next week. If their guides keep climbing — and we expect they will — the shovel sellers keep collecting.

The current phase is simple: hyperscalers are spending ahead of revenue, borrowing to keep building, and every dollar of that $700+ billion in capex lands on a supplier's income statement today. Memory, power, networking — they get paid now. The buyer's re-rating comes later, maybe quarters from now. The supplier's harvest is happening right now.

Korea's panic doesn't stop that. It just put the best-positioned suppliers on sale.

Read more

1. What Changed This Week

1. The bounce confirmed the flush

Tuesday told you everything about who had been selling. Micron jumped 12.2%, Sandisk 14.4%, and SK Hynix’s ADRs 13.8% in a single session, with Western Digital and Seagate up double digits alongside them. The names that led the crash led the recovery. Morgan Stanley called the selloff a “strong entry point,” which is a polite way of saying what we said last week: the sellers were forced, the buyers were not. The trade even has a nickname now. Fortune is calling the three memory giants “Memi,” a $3 trillion sector powering small-caps, emerging markets, and Japan.

And the physical market keeps confirming the squeeze. A mainstream PC memory kit that cost about $75 a year ago now sells for as much as $460 (the same kit, six times the price). The fabs didn’t break. Samsung, SK Hynix and Micron chose to route capacity to AI customers first, and everyone else pays up. That is what pricing power looks like from the inside.

Our posture doesn’t change: ACCUMULATE, in tranches, and if it dips again we keep accumulating. We are not chasing Tuesday’s candle. We are buying a phase of the cycle.

2. Why the rally isn’t done: the buyers haven’t stopped spending

The last cloud cycle had capex following demand at a measured pace. This one doesn’t. Combined hyperscaler capex is headed past $700 billion this year and still accelerating, growing faster than the operating cash flow of the businesses funding it. Free cash flow across the group is compressed. Some of them are borrowing to keep building.

We didn’t have to wait long for proof. Alphabet reported Q2 on Wednesday night: revenue up 24%, Google Cloud up 82% to $24.8 billion, quarterly capex doubled to $44.9 billion. It raised the full-year capex guide to $195–205 billion, from $180–190 billion. It sold $49.6 billion of new stock in June to help fund the build, and free cash flow for the quarter came in at negative $5.9 billion. The stock fell on the print anyway. That one report is the whole setup in miniature: the buyer keeps raising its budget, and the market makes the buyer pay for it. The suppliers collect either way.

Two conclusions fall out of that, and they point in opposite directions for different parts of the tape:

For the hardware suppliers, this is the harvest. As long as the buyers are building, every dollar of that $700 billion lands on somebody’s income statement today: memory, foundry, networking, power. The spend is budgeted and still rising. Our read: that, more than any bounce, is why the memory rally has room left. The rally ends when the spending decelerates. The spending is not decelerating.

For the hyperscalers themselves, the payoff comes later. The compressed cash flow you see in their filings this year is the cost of the build-out, not a warning sign. Our read: the big re-rating moment for this group is still ahead of us, not behind us, and it isn’t here yet. That’s not a reason to sell. It’s the reason the position is early, and why we hold it as one. We’re not in a hurry.

Microsoft, Meta and Amazon report next week into the FOMC, with Apple alongside them. Same question for each: does the capex guide keep climbing? As long as the answer is yes, the shovel sellers’ hour continues.

Read more

1. What Changed This Week

1. Memory crashed 20%+, and the wreckage is leverage, not demand

Micron, Samsung, SK Hynix and the Roundhill Memory ETF are all down more than 20% from their recent closing highs. MU closed at $904 on the 15th, about 25% below the $1,213 record close from June 25. SK Hynix posted its biggest one-day fall in nearly two decades on Monday, down more than 15% as post-Nasdaq-debut profit taking hit. It bounced 13% Wednesday, then fell 11% again today while the KOSPI triggered its 37th sidecar halt of the year.

The interesting part is who was selling. The single-stock leveraged ETFs on Samsung and Hynix that listed in late May are down more than 60% from their June peak, and the biggest of them, a $3.4 billion fund, has lost about 45% since its debut. Then this morning South Korean authorities moved to rein in those leveraged funds, and that announcement alone knocked both stocks down again. When a regulator dismantling leverage is what moves the price, the price is telling you about positioning, not earnings. It didn’t help that Samsung and Hynix had grown to roughly half the KOSPI’s total weight, up from about a quarter at the end of last year. A trade that crowded doesn’t correct politely.

2. While the tape broke, the fundamentals firmed

TSMC reported Q2 today: $40.2 billion in revenue at the top of guidance, 67.7% gross margin above guidance, and a Q3 guide of $44.6 to $45.8 billion, another double-digit sequential step. Capex stays at the high end of the $52 to $56 billion range. ASML raised its 2026 guidance for the second time this year. Samsung’s preliminary quarter was a record, 171 trillion won of revenue at a 52% operating margin. And after the drop, Micron trades at about 6.3 times forward earnings.

One honest caveat, because it’s the real bear case. Morningstar estimates Samsung’s Q2 DRAM price hikes at 30% sequentially versus the 40% it expected, so the pricing slope is flattening. We’ll see the full detail when Samsung reports on July 30. But prices rising more slowly is not a downcycle, and HBM revenue is still tracking from about $35 billion last year to around $60 billion this year.

Our read: the demand curve is intact and the sellers were forced. That flips memory from TRIM to ACCUMULATE. In tranches, though. Hynix bouncing 13% one day and dropping 11% the next says the flush isn’t finished, and we’d rather be early and small than all-in and wrong.

3. BTC cleared the line

Read more

The big story this week is the model race. Grok 4.5 just landed, and it’s good enough to change the math on cost. On the tape, last week’s calls held up: memory rolled over, Apple pushed back toward its highs, and money kept moving into the AI hyperscalers. Two side stories worth flagging. Meta is buying its way into the agent race with a cheap model and a huge user base. And China pulled off its first rocket recovery. Rotation isn’t over. The AI bullrun is still here.

1. What Changed This Week

1. The model race just got hotter, and that keeps capex going

Grok 4.5 shipped on July 8. SpaceXAI trained it alongside Cursor, and Elon’s own read is that it’s roughly on par with Opus 4.7, just much faster and a lot cheaper. Pricing: $2 in and $6 out per million tokens, well under OpenAI and Anthropic (Opus 4.8 runs $5 / $25).

The interesting part is why it got good so fast. SpaceX bought Cursor for $60 billion back in June, so now one side brings the compute and the other brings the data and the real coding workflows. Put those together and you get a top-tier model. GPT-5.5 and China’s GLM are right there too.

So the race is far from settled. If anything it’s getting more crowded, which means capex keeps climbing, not falling. We stay long AI compute and infrastructure. And the core call doesn’t change: memory has peaked for now, and the money is rotating into the hyperscalers and Apple. Micron is down about 22% in two weeks, from its $1,213 record close to around $949. Apple is sitting just under its 52-week high of $317.40, closing at $314.70 on the 9th.

2. Meta: a cheap model plus a huge user base is a real ticket into agents

Meta put out a very cheap model and pointed it at the 3.5 billion-plus people who use its apps every day. The model doesn’t have to be the smartest. If it’s good enough for simpler agent tasks, Meta’s sheer reach can still push it into the main AI conversation. It wins on distribution, not on being the smartest in the room. It also tells you Meta’s AI spending isn’t slowing down. Tencent is the same story in China, a huge ecosystem plus agents, so we’re adding it to the watchlist too.

3. China landed a rocket recovery, and that opens a new space story

On the morning of July 10, the Long March 10A flew for the first time from the Hainan commercial spaceport and reached orbit, and the first stage was caught at sea with a net system. This is China’s first controlled recovery of a launch vehicle, and the first net-capture recovery anywhere. China is now the second country with large reusable-rocket capability.

Our read: on the surface this looks like a threat to SpaceX, but we think it’s a net positive for the whole space and satellite theme. Once both the US and China are committed here, neither one backs off, so spending only goes up. Cheaper launches also change the economics of satellite constellations, turning a money pit into something that can actually scale. That pulls in the whole chain: launch, satellites, ground gear, and comms. We’re going to dig into satellite comms more. For now it stays on the watchlist, not the scorecard.

Read more

1. What Changed This Week

Change 1 · Memory has topped for now

• Micron stalled at $1,250 and rolled over. The print blew past estimates, but the stock still faded on heavy volume. Good news out, no follow through. That’s the cleanest sign of a stage top.

• Money is leaving memory fast. The DRAM ETFs sold off on volume, and SK Hynix and Samsung look the same over in Korea. Foreign investors pulled more than 100 trillion won (about $65B) out of Korea in under two months.

Change 2 · The money is going into the good hyperscalers

• Skip the small-cap noise. The real place for the money leaving memory is the AI hyperscalers. Last Friday chips got hit but GOOG, MSFT and AMZN put in a bottom on volume, and today’s move in META, up on heavy volume, says the same thing.

• The logic behind it is token optimization. As more of the work runs on cheap models for the easy 80% of tasks, the value settles in the layer that charges per token, the cloud and orchestration layer, not the model layer. That’s the hyperscalers’ moat.

• What it means for us: we’re betting on a catch-up move in the hyperscalers.

2. Signal Scorecard

Read more

1. What Changed This Week

Change 1 · Strong dollar, only tech is strong enough

• Warsh’s Fed is hawkish (no cut in 2026, an October hike still live), so the dollar stays bid. A strong dollar drains the two assets that need a weak one: gold broke $4,000 into a bear market (a three-year bull is over) and BTC got pushed back toward $62K.

• Everything that isn’t tech is on the back foot: gold, crypto, EM. Tech is the only real strength left, and even there it’s selective (Change 2).

• This is what a strong dollar does. Real rates are positive, so gold is not a haven and crypto has no liquidity tailwind. Stay short/defensive on both. Don’t fight the drain.

Change 2 · Micron proves the split: spenders fade, earners win

• Micron printed a record $41.46B in revenue, $25.11 adjusted EPS, and guided Q4 to $50B, +15% after hours. But the cleaner signal is who moved with it.

• The memory and equipment names, the ones who get paid for the capex, ripped double digits: MU +15.8%, QCOM +13%, WDC +12.5%, SanDisk +12.3%, Seagate +9.9%, ARM +7.2%, AMAT +7%, ASML +5.6%, INTC +5.4%, AMD +3.8%.

• The hyperscalers who pay for it barely moved or went red: AVGO +2%, NVDA only +0.8%, MSFT −0.0%, AMZN −0.5%, AAPL −0.8%, GOOGL −0.9%.

• That is our whole book on one screen. Long upstream, short downstream. Long hardware, short software. Long AI, short consumer. With the storage cycle re-accelerating, the split keeps going.

Change 3 · SK Hynix’s 7/10 ADR keeps Korea and memory bid

• SK Hynix pulled its US listing forward to 7/10, a ~$29.4B Nasdaq ADR (SKHY) that, at the top of the range, would be the largest ADR ever. Even through Monday’s KOSPI circuit breaker, the memory names led the bounce.

• The listing is a magnet. Expect Korea and the memory complex to stay strong over the next week or two.

Change 4 · Oil erased the war

• Brent gave back every dollar of the war, below the pre-war $72.48 and into contango; WTI broke $70 to a 3.5-month low. Iran keeps de-escalating: more Korean-flagged tankers out of Hormuz, the 60-day sell permit flowing.

2. Last Week’s Calls

Read more

1. What Changed This Week

Change 1 · Warsh’s first FOMC was hawkish

• The Fed held at 3.50–3.75% for the fourth meeting in a row (12-0). The real news was the tone. Warsh didn’t submit his own dot, dropped forward guidance, started five working groups, and opened a review of the $6.6T balance sheet. He also ruled out touching the 2% target.

• The dots moved up. The 2026 median went from 3.4% to 3.875% (one hike implied this year), 2027 from 3.1% to 3.6%. Nine officials now see at least one hike in 2026 (six see two or more), against nine who see a hold or a cut. The committee is split. Inflation got revised up a lot (PCE 2.7% to 3.6%, core 2.7% to 3.3%) and growth got cut (GDP 2.4% to 2.2%).

• Markets took it as hawkish and now price a roughly 70–80% chance of an October hike. S&P −1.2%, Nasdaq Comp −1.3%, 2Y +16bp to 4.21%, dollar +0.7%, gold −1.9% to $4,248, BTC −2.3% to $64,301. The bigger point: Warsh is taking back the “certainty premium” the Fed has handed markets since 2008. Fewer promises, more two-way volatility. It feels like 2021 again. One caveat: this SEP was locked in around the Versailles signing, so the 3.6% PCE dot is probably already stale once you fold in this week’s oil drop (next).

Change 2 · The Versailles deal reopens Hormuz, and oil dumps

• Trump signed the deal at Versailles, effective right away (ahead of the 6/19 date). Hormuz reopens fast, Iranian crude gets an immediate sanctions waiver, and shipping is supposed to be back to pre-war levels within 30 days. Four Iran-linked ships, including two big 2M-barrel tankers, already switched their transponders on and sailed out.

• Brent broke $78, down 15% in four days (its longest losing streak this year), WTI fell to $75.46, and Cushing dropped to a 20M-barrel operating low. Cheaper oil is disinflationary, so it partly offsets the hawkish Fed. That’s why Asia could rally through it.

• Don’t over-read the peace, though. Ballistic missiles were left out, the GOP is angry (Cruz, Cassidy, Graham, Pence), and the swap is lopsided: Iran gets a lot, the US mostly gets back what it had before the war. This is a 60-day ceasefire framework, not the end of the war.

Change 3 · BOJ hikes into the same week

• The BOJ raised rates 25bp to 1.0%, the highest since 1995 (7:1 vote), and said it will stop trimming bond purchases from April 2027, with room for more (OIS sees about 54% for October). The Nikkei briefly traded above 70,000.

• The thing to watch isn’t the hike, it’s the yen. USD/JPY hit 160.75, the weakest since July 2024, even with a hike on the board. Intervention risk is rising, and the carry unwind stays on our watch list.

2. Signal Scorecard

Read more

1. What Changed This Week

Change 1 · AI token demand may be rolling over

• The token spend index — the receipt for what the world actually pays for AI inference — doubled from December, then turned down in recent weeks. Yes, it’s one index from one vendor — but the corporate behavior below points the same way. We addressed this in our SoftBank piece: this index is where the whole $750B capex story starts.

• The demand wasn’t all real. Meta and Amazon shut down their internal token leaderboards, Uber capped AI spend after burning its full-year AI coding budget in 4 months. Part of the usage was a KPI game — and the CFOs are now checking the bill.

• Usage can keep rising while the money falls: spend = volume × price, and price is collapsing. DeepSeek charges 1/200th the price of OpenAI’s flagship, and on the largest model router, ~60% of top-model traffic already runs on Chinese open-source models.

Change 2 · CPI 4.2%

• Headline CPI hit 4.2%, the highest since 2023. Most of it is energy. The war is the inflation. But core was only +0.2%, below expectations, and real wages are falling.

• This doesn’t kill the hike story. It just hands the decision to the Fed. And remember last week’s real lesson: what hurts is the speed of repricing, not the level. One 13bp day did more damage than any fully-priced hike would.

• The next seven days are the densest event window of the year: 6/12 SpaceX debut → 6/15–16 BOJ (a hike to 1.0% is >90% priced) → 6/16–17 FOMC (Warsh’s first meeting). A hawkish surprise burns Japan and Korea. A dovish one burns the dollar — and squeezes Asia’s exporters from the appreciation side. There is no safe outcome — only a choice of what burns first.

Change 3 · SpaceX lists tomorrow

• The biggest IPO ever: ~$75B at a ~$1.8T valuation. 30% is reserved for retail (normal is ~5%). Even crypto retail is rotating into the IPO. 20+ SpaceX ETFs already filed. Schwab client cash is at the lowest since 2019.

• How to read tomorrow: pop or fade → does the buying stick → does it drain the AI-chip complex or bring new money in. Don’t trade the first print — read it.

• Bigger picture: Big Tech has flipped from buying stock back to selling it. Big-five quarterly buybacks are down 74% from the 2021 peak ($12.6B in Q4; NVIDIA’s $80B program is the lone exception), Alphabet is raising $80B+, and ~$4T of valuations (SpaceX, OpenAI, Anthropic) are queueing to list. The bid under US large caps is thinner than the index makes it look.

2. Last Week’s Calls

Read more

The chart is the Silicon Data LLM Token Expenditure Index. Think of it as a meter for how much the world pays for AI inference. It doubled between December and May. In the past few weeks, it rolled over.

Steno’s read: if this keeps falling, the memory, hardware and datacenter trade is over for the cycle.

I think he’s right. But the chart is only the first domino. This piece walks the chain from that one line to the most levered player in the entire AI complex: SoftBank. And it explains why I don’t see a path where the AI trade keeps climbing and SoftBank stays safe at the same time.

Read more

Change 1 · About Broadcom’s “pullback”: a value-capture fight, not an AI-demand bust

• AVGO (Broadcom) sold off after hours, but unpack it and the damage is narrower than the headline: Q3 AI-chip guidance of ~$16.0B came in below the ~$17.2B consensus — a real ~7% miss — and full-year AI chips were guided to $56B vs ~$57.6B expected. But it’s a line-item miss, not a bust: total Q3 revenue was guided above consensus (~$29.4B vs ~$28.6B), Q2 beat (revenue $22.2B, AI semis $10.8B, EPS $2.44 vs $2.39), the $110B backlog held, and CEO Hock Tan left the 2027 outlook unchanged — not cut. Against a stock that had added ~$270B of market cap in the five sessions into the print, an AI line that merely met-not-beat was enough to trigger the unwind.

• The deeper structural shift: the pressure on Broadcom’s AI line is partly a value-capture story: Google is pushing toward more in-house design + a more direct TSMC relationship (the Apple/COT playbook), and custom-silicon competition (Broadcom itself flagged rivals like Marvell) is intensifying — squeezing Broadcom toward the back-end / physical-design layer. The fight is over who captures the capex profit, not capex itself shrinking.

• Little impact on the AI-capex thesis: demand is still being added — Alphabet is raising $80B (incl. $10B from Berkshire) for AI capex, Anthropic keeps expanding TPU capacity, and Apollo / Blackstone are arranging ~$36B of debt to help fund the very chips Broadcom builds for Google. Broadcom’s networking / TPU agreement with Google runs to 2031.

• Broadcom itself: its XPU share / margin to Google is under long-term pressure, but networking / optical (Tomahawk switching, optical DSP, CPO) remains a strong growth engine — expect Broadcom to keep pushing on the optical side. Separately, CrowdStrike fell ~11% on soft Q3 revenue guidance (~$1.21B vs ~$1.23B consensus) — a company-specific guide miss, not a verdict on software broadly.

Change 2 · The mega-IPO drain season begins

• SpaceX’s largest-ever IPO, seeking ~$75B; Anthropic has confidentially filed at a ~$965B valuation; Alphabet is raising $80B for AI.

• The real risk isn’t the $75B raise itself — that number sounds manageable. The danger is overheated sentiment bidding the secondary / aftermarket valuation far too high (SpaceX has been pegged near $1.8T), trapping a pool of capital at the top that is far larger than $75B. The primary supply is just the trigger; the real drain is the market chasing the price with money that dwarfs the raise.

• The danger of a mega-IPO isn’t “how much cash it pulls out,” it’s “at what price, and how many it traps.”

Change 3 · Beyond concentration, a two-sided backdrop

• The credit leg is lighting up: Cliffwater’s private-credit fund faced 17% in redemption requests and gated Q2 at 5%; Partners Group’s Evergreen vehicle saw redemptions spike to 9.8% and also capped them.

• Geopolitics + rates: a lethal US–Iran clash near Hormuz on 6/3 pushed oil higher; US ISM manufacturing rose to 54 (its biggest jump in four years), reinforcing higher-for-longer.

• In last week’s “credit / Fed / geopolitics, two of three” framework, the credit leg is now starting to light.

2. Signal Scorecard

Read more

Read more

From day one of the war, we kept mentioning the duration has not been priced in. Hormuz got locked down by Iran's cheap, asymmetric weapons and the global shipping insurance system. Looking back, the framework held up. Hormuz has now been effectively closed for three months straight. Our call was never "this ends fast." It was "this lingers." And that part has played out.

2. Oil, we got out at the top

We had argued early on that the chokepoint could push oil to a high level, and with hindsight we closed our position almost right at the peak of this leg. The exit, on April 29 to April 30, turned out to be a very good call.

Why didn't it go higher? Everyone started releasing oil from both strategic and commercial reserves, and the US stepped in as the supplier of last resort, which together cushioned the gap. From here, oil might still have room to run, but the risk to reward is no longer attractive. Even if it rises, the hit to the US stays limited: the US is a producer itself, and households in developed economies absorb higher prices more easily. The real pain falls on importers with thin reserves, especially the poorer ones. Economies like India, where the current account and currency are already fragile, take it on the chin first — a case we've leaned on many times in earlier work.

3. Equities and the VIX, the de-risking bias and the great AI decoupling

US households hold ~47% of financial assets in stocks, above the 2000 dot-com peak, so a falling market forces the president to walk policy back. History gives two strike prices: a VIX near 50 triggered last year's tariff climbdown, and a VIX near 30 triggered the dovish Iran turn two months ago. That de-risking bias flips at those thresholds.

AI has made the market numb to oil. Since the late-March ceasefire signals, US equities have decoupled from the shock; the chip rally and better earnings offset the energy hit, and Hormuz faded at the margin. We underestimated this.

This new era is still early. Short term, AI carries real volatility risk — stretched valuations, crowded positioning, a narrative running ahead of itself — so a pullback can come anytime. But zoom out and we're at the beginning: compute supply can't meet demand, and the build cycle for chips, power and data centers has only just started. Short term swings are noise; they don't change the direction.

The paradox is that the stronger the market, the less pressure to force a peace deal. With no market pain, the hawks sound more convincing, and the Hormuz talks keep stalling. Peace now feeds the US mostly through bonds and inflation expectations, not equities, while the hardest-hit oil importers aren't even at the table. The people paying the cost have no voice, and that's what keeps the standoff alive.

4. Gold, not a war hedge but a ticket off the dollar planet

Why did gold pull back after the war? Because war hedging was never the main driver — we laid this out in Day 22 of Hormuz: The First Crack in $39 Trillion. The real driver these past years was central banks: over a thousand tonnes of net purchases a year from 2022 to 2024, led by Poland, Turkey, India and China, motivated by de-dollarization and hedging sanctions risk. That's structural demand, not a war spike.

Gold's real role is a hedge against holding dollars itself. If the dollar ever suffers a deep, sustained devaluation, gold is one of the few tickets off the dollar planet. So it gets supported as the odds of major dollar debasement tick up from zero; there's no second currency ready to take over settlement, and gold is the vehicle for that. Short term it still swings with real rates and the dollar, but that's noise — it doesn't change the long term role.

5. Crypto, the liquidity tide is going out, but respect the cycle

The turning point was last year's 10/10: crypto liquidity dried up and no fresh money came in. Where did the money go? AI stocks became a new "crypto market" of their own, with some names even acting like meme coins, and next to that crypto lost its shine and money got pulled out.

Still, respect the cycle. This is just the old rule playing out — the bull ran too far, so now we're in a bear, and there's no need to be overly pessimistic. The bounces in bear market could still go higher. But don't kid yourself either: that doesn't mean money flows back and we're suddenly in a bull again. The next bull still has to wait for the cycle, so stay patient.

6. The next trade

The market never runs short of opportunities. For capital and individuals alike, the AI era is a once-in-a-lifetime chance. Even a bubble is often the only door an ordinary person gets to walk through: a bubble means dislocated pricing and that mess is exactly what leaves room for the people who come later. If markets were always efficient, maybe only a handful like Buffett would keep making money.

Subscribe

Change 1. Geopolitics and the dollar

• Hormuz. The US hit Iranian targets. Rubio said there’s “good news.” Iran’s state TV leaked a draft deal; the White House called it a “complete fabrication.” Strike, then talk, then strike again. Like we said before, this is not something you fast-forward.

• Long-end pressure. The 30-year held near 5.07–5.18% this week, still at multi-year highs. Yardeni told the Fed in public: drop the easing bias or lose control. The 6/16 FOMC is Warsh’s first real test in the chair.

One catalyst alone still can’t break the market. You need two of three to stack up: credit, the Fed, geopolitics.

Change 2. AI capex spreads from the US to Asia

ByteDance is looking at up to $70 billion in AI capex this year. Tencent and Alibaba are also raising AI spend, though at smaller scale. Analysts now say the spending gap between US and Chinese tech giants is “smaller than the headlines suggest, and smaller still on ambition.” Same week: China tightened travel limits on its top private-sector AI talent. A two-way signal that AI is now a sovereignty story.

Change 3. The scarce-resource satellites are already running

The “scarce-resource satellite” bucket we opened last week (PLTR / SATS / ASTS) is paying off early. ASTS is up more than 50% in a single week. Scarce spectrum, assets you can’t copy, a growth path that does not need AI multiples to keep expanding.

II. Signal scorecard

Read more

The US-China meeting came in mild, pretty much like we expected. Putin’s Beijing visit didn’t flip the table either. But 5/19 was a useful reminder. The 30-year Treasury hit 5.18%, the highest since 2007 — that’s the core signal. Gold and silver and stocks also pulled back in the same window. Then 5/20 Trump softened on Iran, 5/21 AI came roaring back, and the market mean-reverted fast.

The framework hasn’t changed. AI remains the main engine. Chasing up here could still be risky. When a real entry shows up, we’ll tell you. Meanwhile there are alternative instruments.

I. What changed this week

Change 1. Two big events passed quietly

At the airport, Trump got the VP, Han Zheng; Putin got the FM, Wang Yi. On paper Han outranks Wang. In practice the VP slot is largely ceremonial, while Wang sits in the Politburo and actually runs foreign policy. Beijing kept both sides happy without picking a favorite — and Xi’s framing of the Trump visit, a “constructive, strategic and stable relationship”, is exactly the register China wants on the Russia track too.

May 14 in Beijing: the US-China meeting was mild, as expected. No surprise. And no surprise is good news. Then on 5/19 and 5/20 Putin visited Beijing. On energy, arms, or settlement systems, nothing came out of it that would make Washington want to rewrite last week’s meeting.

A reminder, though. A few weeks out, Trump might do something weird like what happened after previous meetings with Xi. It probably won’t come from a direct US-China clash. More likely it comes from the Iran or Strait of Hormuz line. We keep watching.

Change 2. 5/19 was a useful reminder

The core of the day was the long end. The 30-year yield hit 5.18%, the highest since 2007. Gold and silver also pulled back in the same window, which tells you it wasn’t a one-asset story — long-end fiscal pressure was the through-line. Trump was threatening Iran at the same time, so the classic combo was there: stagflation worry, long-end fiscal pressure, and a geopolitical second-order kick. Then 5/20 Trump softened on Iran, 5/21 SoftBank ripped 20%, AI came back, and the tape mean-reverted fast.

Our read is simple. Macro can cause one noisy day, and 5/19 is a clean example, but a single catalyst alone doesn’t break the market. To get real damage you need at least two of three to stack up: credit, the Fed, and geopolitics. We’re not there yet.

Our long-vol tail did its job on 5/19. We’re keeping it on.

Change 3. Warsh is in, and the Fed minutes leaned hawkish

Warsh was confirmed 54 to 45 on 5/13 and took over from Powell on 5/14. His combo is to cut rates, keep QT running, push back on fiscal dominance, and rebuild the line between the Fed and Treasury. That’s not the traditional dove setup. The Apr 28-29 FOMC minutes (released 5/20) also leaned hawkish.

So the tug-of-war between the long end and AI multiples is still going. That’s the main source of noise for the next four to eight weeks. We don’t trade noise.

Change 4. Hormuz is already biting the weaker economies

Hormuz isn’t resolved, but how it spreads is already clear. It hits the weak and the poor first. The all-importers like India and Indonesia are already under pressure. The direct hit to US tech and the US economy is limited. This lines up with our INR short call from before. Position stays as is.

II. Signal scorecard

Read more

Two hundred thousand satellites. One country. One week.

For context: the entire global active satellite fleet today sits at around 15,000 birds (with Starlink alone accounting for roughly 10,000 of them), and the US controls the overwhelming majority of LEO. China just filed for a single-month number that, on paper, would let it own the sky.

Everyone in our world is talking about AI right now, and they should be. AI is the trade of the decade. But sitting right underneath the AI conversation, almost invisible to most public-market investors, there’s a second story building that I think will end up as one of the biggest investment themes of this cycle.

That story is satellites. More precisely, it’s the spectrum and the orbital slots they sit on. They are finite. They are allocated by a UN body almost nobody reads about. And they are increasingly treated, inside every major government, as strategic national assets.

This piece is my attempt to lay out the whole landscape. How the system actually works, who’s fighting whom, which public companies have real exposure, and where I think the money eventually flows.

Read more

The direction has been our base case for weeks: both sides will advance on several fronts — AI, trade, people-to-people exchange, mil-to-mil channels will all get touched — but no single point will see a real breakthrough. What lands on paper will mostly be “framework” and “direction,” not “checklist” and “timeline.”

Net-net, this is a mild, low-risk meeting. Not reconciliation, not loss-of-control. More like both sides resetting the table — installing a “do-not-lose-control” default for US-China relations over the next few years.

What actually matters is not the meeting itself, but the 4-8 week implementation window that begins the moment Trump’s plane lifts off. The specifics — line items, deadlines, deliverables — only land once Treasury, Commerce, USTR, and the White House push out the actual paperwork. Until then, every “super-package” the wires are floating is narrative, not price.

I · What Changed in the Last Week

Change 1 · The US–Iran war window got pushed out, not shut

The Iran line has shifted from “mutual escalation” into a holding pattern — and holding is not over; the safety isn’t off. The cleaner US–China talks resolve, the faster Washington’s bandwidth rotates back to Iran / Hormuz. So the next escalation point most likely shows up on the Iran line. Trump’s own Truth Social cadence during the implementation window is a separate, stacked variable on top of that.

Change 2 · Putin lands in Beijing right after

This is the most under-priced variable this week. As soon as the Xi-Trump handshake settles, Putin arrives in Beijing (timing not officially announced, but the Kremlin confirmed on 5/14 that preparations are complete and the visit will be “very soon”). If the joint communiqué shows visibly deeper Russia–China alignment on any single front — energy, defense procurement, or settlement infrastructure — Washington’s read of “what was that meeting last week” gets rewritten.

Change 3 · The AI Capex top debate hasn’t gone away

DeepSeek V4 — released April 24, running on Huawei Ascend — and Huawei’s CloudMatrix 384 are by now both confirmed in production. AI is also expected to be touched at this week’s talks. Two threads run in parallel: on one side, “China can run frontier models on worse hardware”; on the other, “China is still in the queue for high-end US chips.” Both true simultaneously → NVDA top + China AI supply chain top. Main thesis unchanged.

II · Signal Scorecard

Read more