1October is priced as a coin flip leaning to a hike, and last week's repricing moved the whole 2026 path rather than one meeting
Two venues now agree on the October meeting, and they agree on the awkward side of even money. Polymarket prices a 25bp hike at 53.5% against a hold at 44.5%, on $105k and $115k of 24-hour volume respectively within an event carrying $2.3m of liquidity. Kalshi's October contract last traded at 56 cents for the hike and 45 cents for the hold, with 300,577 and 442,144 contracts of open interest and roughly 28,000 traded on each side in a day. Three points apart, on deep books, pointing the same way.
The more useful number is the week-on-week move. That same Polymarket hike leg is up 17 points over the week, and the hold leg is down 17. The distribution behind it moved with it: the contract on exactly one hike in 2026 fell 35 points in a week to 13.5%, two hikes rose 24 points to 60.5%, and three hikes rose 21 points to 26.4%. A separate contract simply asking whether another hike lands in 2026 sits at 85.5%, on $201k of liquidity.
So the repricing was not about October specifically. The market took last week's decision and moved the entire path up a step, which is why a coin-flip October coexists with an 85% chance of a hike at some point before year-end — a miss in October reads as December, not as a stop.
The decision itself was unusually clean. Alec Renehan, on Equity Mates, reported the hike as unanimous, all twelve voters, taking the target range to 3.75–4.00%, with a majority of participants expecting at least one more before year-end. Renehan also read across to the rest of the world: the ECB raised the week prior, and the RBA meets on 28–29 September with markets pricing a 76–85% chance of a move to 4.60%. This is not a single central bank responding to a single local problem.
What the printed data says about that: headline CPI rose 0.40% in August, putting it near 3.4% year over year, core PCE ran 3.3% through July, payrolls added 162,000 in August, and unemployment sat at 4.1%, down from 4.2% in June. Inflation well above target with a labour market still adding jobs is the ordinary case for tightening, and it is the case the market has now priced twice over.
2The case against raising rates into an oil shock rests on a shock that eased over the weekend
Felix Prehn, on Felix & Friends, made the loudest argument against the decision, and it is a mechanism argument rather than a data one. His claim is that energy-driven inflation does not respond to the policy rate: the price of diesel and fertiliser is set by supply, so raising the cost of borrowing punishes the borrower without touching the cause. He reaches for 1979 — an Iran-driven oil shock, a tightening central bank, and gold moving from roughly $100 to $850. He puts the refinancing arithmetic beside it, citing the Financial Times: about $8 trillion of US government debt rolls in the next twelve months at nearer 5% against an average 3.3% on the maturing stock, which he costs at roughly $130bn a year in added interest, with $457bn of issuance inside the next four days. His conclusion is that rates cannot stay high long enough to work, so the debt gets inflated away instead, and that central banks are already positioned for it — he says their gold buying is running at its fastest pace since about 1997.
The premise is the part that moved. Crude sat above $100 earlier in September; it closed the week at $94.24, down 1.92%, at a more-than-one-week low. Reuters attributed the slide to hopes of progress on Iran diplomacy, and separately reported that China had asked Iran to limit Houthi attacks on Saudi oil facilities. That came after the Houthis claimed strikes on Riyadh, which knocked Saudi and Gulf equities on Sunday, and while vessels were still trickling rather than flowing through the Strait of Hormuz.
The prediction markets took the diplomatic read rather than the escalation read, and they took it hard. Every leg of Polymarket's US-Iran ceasefire ladder rose over 24 hours: 25 September to 92.5% (+6.5), 30 September to 83.5% (+6.0), 31 October to 54.5% (+7.5), 30 November to 42.5% (+6.0), and 31 December to 35.5%, up 12.5 points in a day on a thin $6.6k of volume. That last leg fell by the same 12.5 points on Friday, so the far end of this ladder has now round-tripped inside one weekend and should be read as noise until the deeper legs confirm it. The near legs, which trade twenty times heavier, moved the same direction.
There are two live disagreements here and neither resolves today. Prehn describes a weakening economy being choked further; the payroll and unemployment prints describe an economy still adding jobs, so the two sides are arguing about different economies. And his oil-shock premise is at its weakest point in a week precisely as the decision that responds to it lands. What would settle the first is the next payrolls print. What would settle the second is whether crude holds below $100 through the RBA meeting and into October.
3The binding constraint in the data-centre buildout is the building and its power, not the chip
Two of the weekend's four videos landed on the same constraint from opposite starting points, which is the most interesting thing on the page today.
Brian, on Business with Brian, built his case on physical bottlenecks. A new rack of top-end accelerators draws 140kW and needs under-floor liquid cooling, which he says fewer than 4% of American data centres can accept — he quotes Satya Nadella's line that the problem is not a supply issue of chips but a shortage of "warm shells", finished space with power and cooling already running. Grid connection queues in the biggest markets run around four years. The consequence he draws is counter-intuitive and checkable: old silicon keeps earning, because the alternative to running it is an empty room with a lease on it. He cites signing A100 capacity out to 2029 at pricing above where it sat years ago, and NVIDIA's own CFO saying every chip ever sold and still plugged in is fully utilised, against Jensen Huang's 2025 line that you could not give the prior generation away. He sizes the demand side at roughly $1.7 trillion of already-signed cloud backlog and puts the buildout at 1–2% of GDP in 2026 against housing's 6.6% at the 2005 peak.
Jason Hsu, of Rayliant Global Advisors, on Excess Returns, arrives at the same bottleneck from the rent-capture question. His argument is that a data centre takes five years to build while a model takes a team and a licence fee, so hardware and energy collect most of the rent for perhaps a decade, until over-capacity arrives and the value migrates to whoever integrates cheap infrastructure into applications. He extends it to the US-China race: open-source Chinese models are close enough that model quality converges, and the durable advantage is energy — multi-sourced supply, the largest solar build, nuclear, and a high-efficiency national grid, against a decentralised and old American one. He puts the capability gap at a hair's breadth, noting Stanford has measured it at under 5%.
The two of them disagree about the same capex, and the disagreement is the useful part. Hsu calls a portion of it an unhealthy arms race, where each firm spends because the other is spending and the marginal return falls for both; he is explicit that this will be bad for some of the firms making the investment today, while denying there is any solvency question — these are cash-generative firms pulling future cash flow forward at good rates. Brian's version is that not a dollar of spend has stopped, and he names his own kill-switch: if that $1.7 trillion of signed backlog starts shrinking, his buy signals go with it. One is watching commitments, the other is watching returns on commitments. They can both be right for two years.
4Index concentration is an active bet, and the usual ex-US hedge is being called redundant
Hsu's second argument is about what a passive investor now owns. His framing is that the S&P 500 "has 500 stocks but really it's got seven", that market beta is always a concentrated bet on the theme of the day — internet in 2000, real estate and financials before the financial crisis, AI now — and that a passive allocation today is an active position on one factor held by someone who believes they are diversified. On the recent semiconductor reversal he is unbothered: momentum is a skewed factor, the crashes are what stop the trade being arbitraged away, and the complement is value, which tends to post its best returns exactly when momentum breaks. He also concedes the uncomfortable half — value and small cap have no reliable US premium left, because they are taught in every business school.
Prehn reaches the same worry by a cruder route, estimating that a typical portfolio of a Nasdaq tracker, an S&P tracker and a few technology names carries something like 70% AI exposure, mostly without the holder having chosen it.
The counter came in the same weekend, from Owen Rask on Equity Mates, reviewing a listener's portfolio live. Rask's point is that US-listed companies are already internally diversified — on his estimate 30–40% of the revenue of US firms comes from outside the country — so buying a separate ex-US sleeve to escape US concentration buys less diversification than it appears to. His conclusion for that portfolio was to cut , the ex-US sleeve, as the first simplification, and to keep the global fund that overlaps it. That is a genuine split: two sources say the US technology factor is the risk, one says the standard instrument for hedging it is mostly redundant.
One disclosure belongs with that call, in the same breath. Equity Mates Media is part of the BetaShares group and the episode is sponsored by BetaShares; the sleeve being cut, the fund being kept, and most of the products discussed are BetaShares products. The reasoning stands or falls on its own, but it is not an arm's-length product comparison.
What the sources recommended
Seven single-name calls came out of four videos, and six of the seven come from one channel — which also ran a disclosed paid promotion for an unrelated copper explorer in the same episode, excluded from the tables below because it is advertising rather than a call.
On the buy side, Brian's four all sit inside the data-centre thesis above, and three of the four carry his own hedge. is the unqualified one: he notes it trades at 28 times trailing earnings against a five-year norm nearer 52, with a PEG around 0.35, and says the valuation check returns no problem for the first time on his list. is explicitly sized down — his model reads the price trend as a strong buy but flags that the stock has been this expensive only about 2% of the time in ten years, so he takes a small position and defers the rest to the 30 September earnings report. is cheap on his numbers, under eight times forward earnings on a PEG near 0.3, but has only nineteen months of listed history, so the signal is notched down for want of a track record to test against. is the most conditional of the four: he likes a $25bn backlog against under $1bn of sales this year and $8.6bn of cash, but with four months of listed history his model can only value it off sell-side estimates, and he says a slipped backlog converts him to the sidelines.
Both sells are valuation calls on businesses he describes as performing. gets what he calls one of the strongest sell signals in the video, at roughly 80 times trailing earnings against a 33 times norm, despite Google having taken the right to buy about 6.7% of it against $120bn of chip purchases through 2033. is a partial trim rather than an exit — sales grew 83% and operating margin went from negative to 27 cents on the dollar, but the price sits near 195 times trailing against a norm around 70.
The seventh is from Rask, described above, and it is a portfolio-simplification call rather than a view on the underlying assets.
is discussed at length and carries no usable call: a voice-over in the episode says the information changed between filming and publication, and the rating is left ambiguous.
The week's calendar tests the consumer and the labour market, not the AI trade
Monday is thin — AAR and Ennis , neither a read on anything argued above. The tests arrive from Tuesday. , , , and land together on Tuesday and between them cover the car owner, the homebuyer, the discretionary big-ticket buyer and the builder, which is where a rate path priced for two hikes gets marked against actual demand. and on Wednesday are the labour-market read-through, and they matter more than usual this week because the argument in the second claim above is precisely about whether the economy is weakening. , and close the week on Thursday.
The AI claims do not get tested this week. The dated test for the memory shortage is on 30 September, which is also the report Brian said would decide whether he buys the rest of his position. Before that, the RBA decides on 28–29 September, and the UN General Assembly runs alongside a Trump-Xi meeting that the Treasury Secretary has already trailed as following successful talks — a market prices a US-China tariff agreement by 31 December at 88.2%, though on only $3.8k of 24-hour volume, which is too thin to lean on.
