Intelligence as Infrastructure: How AI Is Rewiring the Economy

Summary

• Leon’s macro sequencing: AI is “long-term… a highly deflationary force” — healthcare alone is ~18% of GDP and ripe for LLM disruption — but the short-term is sticky inflation first: chip/memory/infrastructure input costs are skyrocketing and software-engineer hiring rose 18% last month as the old economy staffs up to implement LLMs. Add robotics and “the labor market can look very different a few years from now” — “I do not envy” the Fed.
• Alex’s adoption math is the core bull case: only ~10 basis points of 1 billion white-collar workers use AI in the advanced agentic way, and Claude Code has just 14 million DAUs — headed to 500 million — while those power users burn “a thousand times as much compute.” This isn’t S-curve adoption, “this is an L-curve, straight up,” and “we have half of what we need from a compute standpoint before it’s even started.”
• Alex’s foundation-model economics put the capex-ROI debate to bed: Anthropic has gone 100M → 1B → 9B → $45B run rate, Anthropic + Open AI could hit “$200 billion of revenue” by year-end, and because they locked up compute early, incremental margins are enormous — “you could be looking at something that’s like 18 times earnings.”
• Leon’s boldest sector call: WFE spend goes from ~$120-130B to $300B over 3-4 years as the industry shifts from one under-spending buyer (Taiwan Semi) to multiple spenders (Intel, Samsung, Hynix, Micron, SanDisk), with customer margins at 70-80% and LTAs reducing cyclicality. “Estimates are like 50 to 70% too low… I don’t know what the next 10% is, but I’m guessing the next 50 to 100 is up.”
• Alex’s “golden age of hardware”: 40 years of commoditized $2,000 x86 servers are giving way to $300,000 racks that must be reinvented yearly — networking jumping 400G → 800G → 1.6T → 3.2T — yielding units +50%, ASPs +20-100%, gross margins +300-500bps, and “growing earnings 100% for the next 4 years.” Efficiency gains won’t kill the trade: tokens grow ~14x/year vs. chips improving 100-200%.
• Picks: Alex owns TTMI (complex PCBs going from 10 to 120 layers, just won Nvidia, 40% defense incl. Iron Dome) and Google (“they’ve won AI… could easily be up 50%, I don’t see very much downside”). [Speaker?] likes Lam Research ($55B revenue potential on the memory boom, a mid-‘27/‘28 story) and tight analog — likely Infineon, Texas Instruments, likely Renesas.
• On software, both are cautious but split by vertical: horizontal application layer has “a lot of trouble,” but data/infrastructure names like Datadog benefit — Anthropic uses its tools. Alex: the decline is “largely justified”; AI now tops the CIO’s list and token spend is taking software budget.

Deep dive

1. Sticky inflation first, deep deflation later — and a Fed with no good tools

• Leon’s frame is inflation causation: long-term AI is “a highly deflationary force” — “we’re going to get a lot more for a lot less,” with healthcare (~18% of GDP, 10% of individual income) his example of what LLMs disrupt. But there’s a time lag: near-term, CPU/memory/infrastructure pricing “is skyrocketing” as an input cost.
• The counterintuitive labor datapoint: software engineers — the job you’d expect AI to kill — saw an 18% hiring increase last month. Leon’s read: it’s not tech (they’re cutting) but the old economy hiring engineers and product developers to implement LLMs. As prompt engineering advances, “there may not be so much need for these guys” — and with robotics layered on, “I do not envy” the Fed over the next few years.

2. We’re at 10 basis points of adoption — this is an L-curve, not an S-curve

• Alex’s staging: everything so far is “AI 1.0… search engine on steroids.” Business AI is “Claude Code or something like that plugged into all your data sources,” with skills and agents on top — and maybe 10bps of 1 billion white-collar workers use it that way. Whale Rock itself is “looking to hire Claude ninjas.”
• The compute implication: Claude Code has only 14 million DAUs, headed to 500 million, and the true power users burn “a thousand times as much compute and tokens” as everyone else. “We talk about S-curve adoption. This is an L-curve, straight up” — and “we have half of what we need from a compute standpoint right now before it’s even started.”

3. Foundation-model economics end the capex-ROI debate

• Alex’s comparison from experience — his first stock at Fidelity was Amazon in ‘98, when there were 100M internet users and 2M e-commerce users: “This one’s moving faster.” Anthropic’s revenue went 100M → 1B → 9B → $45B run rate, maybe $100B. In the AI stack — chips, clouds, foundation models, applications — the two value-capture layers are foundation models (Whale Rock owns Google, Open AI, Anthropic) and chips.
• Alex’s margin kicker: Anthropic + Open AI could reach “$200 billion of revenue” toward year-end, and because they locked up compute for years while token pricing rises, incremental margins are enormous — “you could be looking at something that’s like 18 times earnings. So that argument would be put to bed.”

4. Semi equipment: from one spender to many — WFE to $300B

• Leon’s setup: demand rolled from GPUs to memory to CPUs and networking, creating “massive supply constraint,” while a decade of boom-bust left the industry disciplined — really one big spender, Taiwan Semi, which by profitability-over-capex metrics “underspent significantly” and, given Intel’s and Samsung’s foundry wins, “maybe made a mistake.”
• The call: multiple spenders now (Hynix, Micron, SanDisk — “what’s the last time we talked about a NAND cycle? Must be a decade ago”), so WFE goes from ~$120-130B to a “$300B mark over the next three to four years.” Customers carry 70-80% margins vs. semi equipment’s 50%, LTAs give multi-year visibility and less cyclicality, and “estimates are like 50 to 70% too low.” Hedged as hedged: “there’s certain things that have gone 100 miles an hour in a 60 mile an hour zone, so there’s going to be some accidents… I don’t know what the next 10% is… but I’m guessing the next 50 to 100 is up.”

5. The golden age of hardware — and why efficiency won’t save you compute

• Alex’s history: for 40 years hardware was a commoditized $2,000 x86 server, compute demand grew 30% and Moore’s Law matched it. Then AI — Elon’s “supersonic tsunami,” 10x-ing yearly — forced innovation at every layer of $300,000 racks: PCBs only 2-3 companies can make, networking jumping 400G → 800G → 1.6T → 3.2T annually. The algorithm: units +50%, ASPs +20-100%, margins +300-500bps, earnings +100% for 4 years — “the multiples haven’t caught up. It’s all been earnings.”
• The host’s efficiency challenge, and Alex’s answer: tokens are growing ~14x a year while chips improve 100-200%, so even 2-3x efficiency gains “can’t keep up” with token demand.

6. Software bifurcates; the picks are TTMI, Google, Lam, and tight analog

• Both reject “software” as one trade. Leon: horizontal application layer has “a lot of trouble,” but data-driven/infrastructure software wins — Datadog is “a unique asset… at the heart of actually benefiting.” Alex is harsher: old code is “horse and buggy,” the new way is “the transporter from Star Trek”; AI tops the CIO list, tokens eat software budget, and software firms’ own AI products have “been kind of a fail so far” — though Anthropic using Datadog’s tools is “a pretty good tell.”
• Alex’s picks: TTMI — PCBs going from 10 layers to “20, 30, 40, even 120,” just won Nvidia, 40% defense including Iron Dome — and Google: “They’ve won AI… the only public company with a foundational model,” TPUs now powering Anthropic, “could easily be up 50%. I don’t see very much downside.” [Speaker?]’s: Lam Research — memory-heavy, a “middle of ‘27-‘28 story,” $55B revenue potential with the street 50-70% too low — plus analog semis that “could look like memory from a pricing standpoint”: likely Infineon (pitched earlier at the conference), Texas Instruments, likely Renesas. And the kicker: [Speaker?] bought his mom the SMH — “she’s going to keep holding it after what Leon said.”

Verification Notes

• The raw captions do not directly identify the speaker who says they bought SMH for their mother; “[Speaker?]” is retained.