The Fourth Leg: Where the AI Trade Goes After Memory

A note on dates: this framework was developed on June 10, 2026, and the market data in it is deliberately preserved with its June 2026 as-of dates rather than refreshed, because the point of publishing it now is to show the framework as it stood and what it anticipated. Where late-August developments bear on the thesis, they appear in their own section, dated.

The Framework: Constraint Migration

Each wave of the AI buildout has followed the same sequence: a constraint emerges in supply-chain checks, earnings revisions outrun multiples for the companies that own the constraint, a narrative forms, sell-side coverage saturates, retail and systematic flows arrive, and the trade becomes consensus. The investable window, in our view, sits between the first evidence of the constraint and the formation of the narrative. Here is how the framework has performed as a process, including the failure:

The Miss, Owned

Memory re-rated from late 2025 into what Bank of America's May 2026 Global Fund Manager Survey, per press summaries, called the most crowded trade in markets: 73% of managers named long global semiconductors the most crowded position, which those summaries described as the highest reading in the survey's modern history. Micron reached a market capitalization of roughly $1.05 trillion and SanDisk was up roughly 465% year to date, per market trackers, as of early June 2026. We watched the same supply-chain checks everyone else did, and we did not act on them before the move. There is no version of this note that is honest without that paragraph.

Two lessons from the miss govern everything below. First, the window is compressing: memory took roughly two quarters to travel from first HBM evidence to full crowding, and the 800VDC power narrative pre-priced its most obvious pure play within weeks of partnership announcements. Second, the migration is running out of hardware. Compute, generation, memory, storage, optics, cooling, packaging equipment, and turbines have each been discovered in turn. The two largest constraints we believe remain unpriced are the conversion of capital spending into revenue, which lives in software, and the last physical layer the market treats as a commodity: power-conversion silicon inside the rack.

What the Market Had Already Found (as of June 2026)

Before naming the next leg, the framework audits what is already priced. The snapshot below uses June 2026 data; performance figures are approximate and drawn from market trackers, with secondary sources flagged in the Sources section. This is where the AI trade stood on June 10, 2026:

The stress test arrived the week the framework was written. On June 3, 2026, Broadcom guided third-quarter AI revenue to $16.0 billion against a consensus near $17.2 billion, per press reports of the call; on June 5 the Nasdaq fell 4.2%, its worst day since April 2025, with the damage concentrated in the crowded layers: Micron fell roughly 20% over two sessions and roughly $1.3 trillion of chip market capitalization was erased in two days, per a secondary press estimate. In the same May 2026 survey cited above, 34% of managers named AI capital spending the most likely source of the next systemic credit event. Our conclusion in June, which we still hold as opinion: incremental dollars into consensus AI hardware were buying positioning risk, not insight, and the next leg had to come from outside that table.

The Fourth Leg: Agentic Software, the Monetization Trade

The combined capital spending of the largest hyperscalers was running near $725 billion for 2026, up roughly 77% year over year per company guidance and sell-side aggregations as of June 2026, with street estimates above $1 trillion for 2027. Our view: that spending only makes sense if AI revenue shows up somewhere, and in the first half of 2026 it began showing up inside the software companies the market had spent the first quarter pricing for obsolescence.

The "seat compression" fear, that agents replace knowledge workers and therefore per-seat software dies, produced the worst software quarter since 2008: the IGV software index fell 24% in the first quarter of 2026, per index data, with roughly $2 trillion of software market value erased in about thirty days, per press aggregations, and large platform names down more than 30%. At the same time, the disclosed numbers turned. The evidence, all from primary company reports:

Salesforce (February 25, 2026 results, fiscal Q4 2026): Agentforce annual recurring revenue of $800 million, up 169% year over year; Agentforce plus Data 360 ARR above $2.9 billion, up more than 200%, a figure that includes roughly $1.1 billion from the Informatica acquisition. As of early June 2026 the stock traded near 13 times forward earnings, per market data, a multiple that in our reading priced secular decline against an accelerating disclosed agent revenue line.

Snowflake (May 27, 2026 results, fiscal Q1 2027): product revenue of $1.33 billion, up 34% and accelerating, raised guidance, and a disclosed commitment of up to $6 billion with AWS. The next session the stock rose 36%, which press reports described as its largest single-day gain on record. We cite that session for a specific reason: it demonstrated, live, the mechanism this thesis depends on, a hated complex re-rating violently on agent-revenue proof.

Palantir (May 4, 2026 results, Q1 2026): revenue of $1.63 billion, up 85% year over year, which the company noted was its fastest growth since 2020, with US commercial revenue up 133%. The stock was nonetheless down roughly 20% year to date at the time of the print, per market data.

Two more datapoints framed the disconnect. Deutsche Bank reported, per secondary press coverage, that it could not find a software company in its survey work expecting a negative 2026 revenue effect from AI. And roughly 70-80% of IGV constituents beat the quarter the market was busy selling, per press aggregations. All figures above are as of June 10, 2026.

Why Software Is the Memory Analog

The reason memory worked, in our reading, was not that DRAM was secretly a great business. It was that the market held a settled, wrong narrative ("commodity cyclical, structurally oversupplied") at the exact moment AI structurally changed the demand curve, and the whole complex was priced off the old narrative. The structural mapping, as we drew it in June 2026:

Two honest differences, both of which cut against overexcitement. First, the software case is a multiple-recovery story layered on solid earnings growth, not memory's earnings explosion; the payoff profile is meaningfully more modest, and anyone mapping memory's magnitude onto software is misreading the analogy. Second, seat compression is not imaginary. Some per-seat models will genuinely shrink. The version of this thesis we hold therefore favors companies that price on consumption, data, or outcomes, or that have already converted agents into a disclosed and accelerating revenue line, over undifferentiated per-seat tools. That is a view about business models, stated as opinion, not a recommendation about any security.

One more property of this leg matters to how we think about the whole AI complex: it is where the central question gets answered. If agent revenue keeps inflecting, the monetization story that justifies more than $1.7 trillion of 2026-2027 hyperscaler capital spending, on street estimates, is real. If it stalls, that spending, and every crowded hardware name attached to it, is mispriced. Watching the software complex is watching the evidence on the question the entire AI trade depends on.

The Physical Continuation: 800VDC Power Silicon

The runner-up theme in the June framework is the purest continuation of the physical-bottleneck playbook: the power constraint our generation work identified at the site level is migrating inside the building. As AI racks scale from roughly 120 kilowatts in the Blackwell generation to roughly 190 kilowatts with Vera Rubin (ramping as of mid-2026, per NVIDIA disclosures), toward 600 kilowatts with Rubin Ultra in 2027 and 1 megawatt beyond that, NVIDIA has pushed the industry toward an 800-volt DC power architecture, announced in May 2025 with full production targeted alongside its Kyber rack generation in 2027, per the company's developer publications and Computex 2026 materials.

The content math is the striking part, and it comes with a flag. Morgan Stanley has estimated that power-semiconductor content per rack rises from roughly $11,200 in the B200 generation to roughly $191,000 by the 2028 Feynman generation, about seventeenfold in three years. That estimate reached us through secondary reporting and is, to our knowledge, a single analyst's work; we treat the direction and slope as the signal, corroborated by NVIDIA's own architecture disclosures, and we do not treat the dollar figures as independently verified.

The complication, and the reason this ranked second rather than first: the narrative already existed by June 2026. The most obvious pure play, Navitas, traded 35-40% above consensus price targets on roughly $45 million of trailing revenue, per market data at the time, and the premium franchises (Monolithic Power near 60 times forward earnings, Vertiv near 42 times, per market data) were priced as such. What looked unpriced to us was the lagging expressions: onsemi, whose AI data-center revenue roughly doubled year over year in the first quarter of 2026 and was guided to roughly double again, per company reports, while the stock was valued like an automotive cyclical; Infineon, which guided AI data-center revenue from roughly 1.5 billion euros in fiscal 2026 to roughly 2.5 billion in fiscal 2027 and raised capital spending by 500 million euros to expand AI power capacity, per company disclosures; and Flex, which announced on May 5, 2026 the spin-off of its Cloud and Power Infrastructure segment into an independent public company targeted for the first quarter of 2027, a hard event that forces the market to value that business against power-infrastructure comparables. Those are observations about where the theme's exposure sat cheaply in June 2026, stated as opinion, not recommendations.

Further Down the List: Three Themes We Ranked but Did Not Lead With

Grid materials. Every 2027-dated supply-chain check we reviewed names power availability as the binding constraint on the AI buildout; ERCOT raised its 2030 data-center demand forecast from 29 to 77 gigawatts in a single planning cycle, per its published planning documents, and US transformer lead times have been reported above 80 weeks, per secondary industry reporting. The equipment layer is fully discovered; the quieter corner is grain-oriented electrical steel, where Cleveland-Cliffs describes itself as the only domestic producer and opened a distribution-transformer plant in Weirton, West Virginia in 2026, per company releases. This is a special situation wrapped in steel-cycle risk, not a sector call.

Advanced packaging and test. The physical bottleneck is real (TSMC's CoWoS advanced packaging has been reported sold out with 50-plus-week lead times, per TrendForce and DigiTimes reporting, which we flag as secondary), but the equity discovery largely happened in 2025-26, and hybrid bonding, the technology change bulls lean on, is not required for the HBM4 generation, per TrendForce reporting on SK Hynix's process choices and JEDEC standards discussions. We saw relative value only in inspection, metrology, and outsourced-assembly laggards, and only on broad flush days.

AI for science. The one genuinely pre-narrative theme on the list: as of mid-2026 no AI-discovered drug had received FDA approval, to our knowledge, and the public pure plays sat in multi-year drawdowns. But the catalysts now carry dates: Insilico Medicine received FDA clearance for an AI-designed Parkinson's candidate in January 2026, per its release, and Alphabet's Isomorphic Labs raised a $2.1 billion Series B in May 2026 with first-in-human trials targeted around the end of 2026, per its funding announcement. This is venture-style optionality with total-loss outcomes among the possibilities, and we treat it that way.

What Late August Has Confirmed So Far

This note is being published eleven weeks after the framework was written, which allows one dated check against subsequent events. On August 26, 2026, NVIDIA reported second-quarter fiscal 2027 results and guided gross margin to a trough of 71-72%, attributing the pressure partly to memory costs, per the company's guidance on the call. We read that as a confirmation of the framework's core mechanic: the memory constraint that re-rated the memory makers is now visible in the cost structure of memory's largest customer. Constraints migrate, and their costs migrate with them. We would caution against reading more than that into a single print, and we note that the agentic-software and 800VDC legs of the thesis will be proven or broken by their own disclosures over the coming quarters, not by this one.

The Bear Case, Steelmanned

The strongest objections to this framework, stated as forcefully as we can state them:

The rate regime. As of the June 10, 2026 writing, consensus expected May CPI near 4.2% year over year, the hottest reading since early 2023, and market-implied odds of a December rate hike were near 70% after a strong May payrolls report, per market pricing at the time. A genuine hiking cycle compresses every long-duration multiple in this note, software first. The monetization thesis can be right on fundamentals and still lose for a year on discount rates.

Seat compression proves broad. Our top theme asserts that the first-quarter panic overpriced a real but narrow phenomenon. If renewal cohorts through fiscal 2027 show measurable seat loss at the large platforms, not just at undifferentiated point tools, then the market's Q1 verdict was early rather than wrong, and the software leg fails.

An AI-financing credit event. Circular financing arrangements, in which chip and cloud vendors help fund the labs that buy their output, were estimated above $800 billion as of mid-2026, per secondary press estimates, and 34% of surveyed fund managers named AI capital spending the likeliest source of the next systemic credit event, per press summaries of the May 2026 BofA survey. A funding accident would hit every theme in this note. We would argue software is the most insulated leg, since its revenue comes from customer operating budgets rather than project finance, but insulation is not immunity.

Monetization lands somewhere else. Agent revenue could concentrate in the hyperscalers and model labs rather than in application software vendors, in which case the capital spending monetizes but the software complex never re-rates. The disclosed agent revenue lines at the application layer are the evidence against this so far; they are also young.

The 800VDC curve is one bank's spreadsheet. The seventeenfold content estimate is a single analyst's work reaching us through secondary reporting. The architectural direction is corroborated by NVIDIA's own publications, but the dollar magnitudes could prove substantially overstated, hyperscalers could retain existing power distribution outside NVIDIA reference designs, and rack-generation timelines slip routinely.

Crowding contagion. With semiconductor positioning at record crowding per the survey data above, forced de-grossing in AI hardware spills into everything AI-adjacent on stress days; June 5, 2026 took even cheaply valued power-silicon names down double digits. Uncrowded legs offer no drawdown immunity, only, in our view, faster recovery. Anyone buying this theme expecting it to decouple on the bad days will be disappointed.

What Would Prove Us Wrong

We publish break conditions on every framework we maintain. For the Fourth Leg, the conditions are qualitative and observable:

1. Agent revenue decelerating hard. If the disclosed agent revenue lines at the large software platforms decelerate sharply and persistently from their 2026 growth rates, the monetization inflection this framework rests on is not real, and the software leg fails on its own evidence.

2. Seat loss showing up in retention. If net revenue retention at the anchor platforms begins visibly reflecting seat compression across multiple quarters, the bear narrative the market priced in the first quarter of 2026 was correct, and we were the ones fighting the tape.

3. Monetization skipping the application layer. If, over the coming disclosure cycles, AI revenue growth concentrates overwhelmingly in hyperscalers and model providers while application-layer agent lines stagnate, the fourth leg exists but we picked the wrong expression of it.

4. 800VDC failing to leave the reference design. If major hyperscalers decline to adopt 800-volt DC distribution outside NVIDIA's own rack architectures, or the next rack generations slip materially, the power-silicon content story loses its slope and the runner-up theme reverts to an ordinary cyclical.

5. The macro overwhelming everything. If a sustained rate-hiking cycle or an AI-financing credit event arrives, every leg of this framework de-rates together regardless of fundamentals. That would not disprove the constraint-migration logic, but it would make the framework unactionable for as long as the regime lasts, and we would say so rather than pretend otherwise.

As of this writing, none of the five is in evidence. That statement is a snapshot, not a promise, and we revisit it as the disclosures arrive.

Our View, Stated as Opinion

Three Douglas Research believes the AI trade's next durable leg is monetization, best evidenced in agentic software, with 800-volt DC power silicon as the physical continuation of the constraint-migration sequence, and we believe the crowded hardware expressions of the AI trade now carry positioning risk that the neglected legs do not. Those are opinions, formed on June 10, 2026, published with their original as-of dates, and falsifiable by the break conditions above. We were early twice and wrong once inside this same framework, and we think the miss makes the process more credible, not less: it is why every claim in this note carries a date and a source, and why the bear case is printed at full strength. Our company-level work sits behind the membership wall.