Actionable insights — Bottleneck rotation and the winner-count test

0:00 1. The bottleneck-rotation screen — buy the scarce link, and expect it to move

1. Start from the spend, not the story: establish where the capital is physically going. Here, "into infrastructure which goes right back to chips" — the money ends up in hard components, so the analysis belongs at the component level.
2. Map the chain end to end (accelerators → CPUs → memory → networking → packaging → power → land and construction) and ask, at each link, which one cannot be expanded fast enough right now. That link is the bottleneck and it holds the pricing power.
3. Confirm the shortage is durable rather than a single quarter's squeeze — "there's shortages in there that are going to continue for the next couple of years" — because a bottleneck that clears in one quarter never pays.
4. Use relative performance inside the theme as the diagnostic, not the conclusion. When two names on the same chain diverge violently, the market is telling you where the constraint has moved. Read the spread; don't chase it blindly.
5. Hold the position as a bet on the constraint, not the company, and set the exit accordingly: when supply catches up at that link, the reason to own it is gone even if the company is fine. "The bottlenecks are going to keep switching."
6. Then re-run step 2 to find the next scarce link before consensus does — the framework's value is in the rotation, not in any single name.

• Lead times and spot prices at each link (CPUs, HBM/DRAM, substrates, optics, transformers, turbines); capacity announcements at the current bottleneck — the signal that the rotation is coming; the previous bottleneck's pricing rolling over; a name still being bought for a shortage that has already cleared.

0:47 2. Count the winners a category supports — before analysing any single company

1. Before valuing anyone, ask the structural question: how many durable winners does a category of this shape historically support? "You only have a certain number of winners."
2. Answer it from base rates in analogous platform markets, out loud: "who's the winner in search? That's just Google. What about in e-commerce? That's just Amazon." Both converged on one.
3. Use that count as a hard ceiling and apply it to the current theme: "are you going to have five different guys win in AI? No." If the market is pricing more winners than the category has ever supported, someone in the group is mispriced downward-bound.
4. If the count is greater than one, find the axis the winners split along — usually customer type, not technology. Here it is consumer versus corporate.
5. Assign each slot on a specific, checkable advantage, not on brand: consumer went to the owner of the complete stack (chips → data centres → models → distribution); corporate went to the firm with the demonstrated financial evidence — profitability plus an unprecedented revenue ramp.
6. Whatever is left over after the slots are assigned is your structural short/avoid, by construction rather than by criticism — "stuck between the two of them… jammed between the other two guys."

• Category economics that argue for more than one winner (real switching, multi-sourcing, commoditised inputs) — that breaks the base rate; the squeezed player buying a distribution slot (a device, an OS, a browser deal) to escape the middle; a stack owner losing a layer; the corporate winner's profitability reversing, which was the whole reason for the slot.

1:11 3. Rank private AI companies on two hard financial facts, not on model quality

1. When asked to rank companies you cannot value from a screen — private labs, pre-IPO names — refuse benchmark scores and product buzz as inputs; they change monthly and are marketed.
2. Demand two things that are hard to manufacture: a crossing into profitability (the firm takes in more than it spends, despite the cost of training) and a revenue ramp abnormal for the company's size — "ramping like nothing we've ever seen in history for a company of that size."
3. Frame the choice as exclusive. Asked whether both private labs interest him "as stocks on the public market," he names exactly one — a forced ranking is more informative than two favourable opinions.
4. Carry the same test into the eventual listing: the one that already funds itself needs the IPO less, which is precisely why it is the better one to want.

• Profitability that depends on a single large partner or on capitalised compute; growth rates decelerating as the base grows; a switch from "profitable" to "profitable before training costs"; funding rounds at flat or down valuations for the squeezed player.