Part IV — Where It's All Going

This is the forward view, and it is an argument, not a list. From what the evidence shows — the roadmaps with dates, the capacity forecasts, the named projections of Goldman, Morgan Stanley, JPMorgan, the IEA, Epoch, and METR — it reasons toward what happens next and, crucially, where the opportunity sits. The through-line is a single moving target: the binding constraint keeps migrating, and the money follows it.

4.1 The master pattern: the constraint keeps migrating

The most useful thing to understand about the next two years is that the scarce input in AI is not fixed. In 2023 it was the GPU; by 2026 it is memory, packaging, and above all power; and each time the bottleneck moves, the pricing power — and the best trade — moves with it. The evidence for the migration is concrete. On the compute side, Nvidia has more than $1 trillion of Blackwell-and-Rubin purchase orders visibility through 2027 (Jensen Huang, GTC 2026), with Blackwell sold out into late 2026, so supply, not demand, is the binding constraint there. One layer down, TrendForce projects high-bandwidth memory rising from 18% of DRAM wafer input at end-2025 to 30% by end-2027, with HBM contract prices possibly doubling in 2027 — a crowd-out that hands the memory makers pricing power. And one layer down again, transformer lead-times run three to four years, and roughly 7 of 12 gigawatts of US data centers planned for 2026 have already been canceled or delayed for lack of power and gear.

That last number is the tell. When announced compute gets canceled not for lack of chips but for lack of transformers, the constraint has demonstrably left the chip. The investable conclusion that runs through everything below: own whatever the constraint currently gates. Today that is power and the electrical supply chain; a year ago it was packaging; and the discipline is to keep asking where the scarcity is, not where it was.

4.2 Compute: the value migrates off the die

Nvidia's own roadmap tells the story of where compute value is going. Vera Rubin ships in 2026, Rubin Ultra in the second half of 2027 (15 exaFLOPS of FP4 inference, 384 GB of HBM per GPU, per TrendForce), and Feynman in 2028 — and Feynman's defining feature is custom HBM and 3D die-stacking, not a faster logic core. The architecture itself is telling you that the differentiation is migrating off the logic die into memory and packaging.

The competitive evidence points the same way. Counterpoint and JPMorgan project custom ASIC shipments surpass merchant GPUs in units in 2027 (roughly 12.5M ASIC vs 10.9M GPU of 23.3M total), with Broadcom holding about 60% of AI-server ASIC design share. The reasoning that matters for an investor is that units are not revenue and neither is computing power: Nvidia keeps the high-margin frontier-training and merchant markets while ASICs absorb captive, stable inference, so Nvidia's dollars can keep rising even as its unit share falls — the pie is growing fast enough (inference is already two-thirds of compute, and Gartner sees AI-IaaS inference spend at 65%+ by 2029) that both can be true.

The opportunity, therefore, is not limited to picking the GPU-versus-ASIC winner; it also includes the layers into which system value is migrating. Three merit particular attention. Memory is a direct but cyclical expression: HBM growth can improve mix while crowding conventional DRAM wafer capacity, but high margins invite supply response. Networking and optics are the second: Broadcom spans custom ASICs, AI Ethernet, and co-packaged optics, alongside more concentrated component suppliers with greater customer risk. Advanced packaging is the third and a near-term deployment constraint. These are strategic observations; entry valuation and capacity normalization determine the eventual security return.

4.3 Models: capability compounds, ROI lags, and the variable that reconciles them

The models layer presents the sharpest tension in the whole atlas, and it is worth stating as a genuine contradiction. On one side, capability is compounding faster than almost anything in technology: Epoch AI measures frontier training compute growing about fivefold a year (and expects that to continue through 2030), with algorithmic efficiency improving threefold a year on top; METR finds the length of task an AI agent can complete reliably is doubling roughly every three months and accelerating. On the other side, MIT found that about 95% of enterprise AI pilots delivered no measurable profit. Capability is racing; deployed value is lagging.

The reconciling variable — and the thing to watch above all in this layer — is agent reliability at long horizons. The METR curve says multi-hour, then multi-day, tasks become reliably automatable within a couple of years; the moment that reliability crosses the threshold enterprises need, the MIT failure rate inverts and the "GenAI divide" closes. Until then, the demand engine is not the pilots but the token explosion: Goldman Sachs projects token demand growing roughly 24-fold to 2030 as inference cost falls 60–70% a year and agentic workflows consume 5–30× the tokens of a chat query. That is a Jevons dynamic — cheaper inference expands, not shrinks, total consumption — and it is why aggregate AI spending (Gartner: $2.5 trillion in 2026, rising to $3.3 trillion in 2027) keeps climbing even as per-token economics collapse.¹

The opportunities that follow. First, the efficiency curve favors the inference-infrastructure and agent-orchestration layers over the frontier-model layer, because the value accrues to whoever serves the exploding token volume, not to whoever trains the marginally-best model that open weights will match within a release cycle. Second, physical AI is the next compute S-curve: Nvidia's Cosmos world model, Google's Genie, and the humanoid wave (Goldman sizes the humanoid market at $38 billion by 2035; Morgan Stanley, far more aggressively, at $5 trillion by 2050) shift demand toward simulation and robotics compute and toward the humanoid supply chain of actuators, sensors, and batteries. Third, the open-weight convergence on coding compresses the margins of the closed coding models even as it expands the deployable base — a reason to favor the compute and the self-hosting infrastructure over the model vendors. And if Epoch's warned-of compute crunch materializes for long-context agentic workloads, pricing power swings up the stack to the compute owners and away from the application startups.

4.4 Power: the physical deployment constraint

Nowhere is the "own the constraint" logic more actionable than power, and the forecasts are unusually aligned on direction. The IEA sees data-center electricity roughly doubling from 415 TWh in 2024 to about 945 TWh by 2030; Goldman Sachs sees global data-center power up 165% by 2030 and US demand doubling from 31 GW in 2025 to 66 GW by 2027; Morgan Stanley sees a roughly 49-gigawatt US shortfall and data centers reaching ~18% of US electricity by 2030. The disagreement is only about the severity of the shortfall, not its existence.²

The reasoning that turns this into a trade is that the binding constraint is not generation in the abstract but the equipment and the lead-times: transformers at three-to-four years, gas turbines sold out through 2030 (GE Vernova's backlog reached 116 GW), and the 7 GW of canceled 2026 capacity as the first hard proof. The opportunities cascade from that. The electrical-equipment and turbine makers — Eaton, Siemens Energy, GE Vernova, Hitachi Energy, Quanta — have multi-year backlog visibility and pricing power, which is why they have already re-rated hard (a caution: much of the good news is priced, so the edge is now in the lead-time and book-to-bill data). The behind-the-meter bridge is the breakout second-order winner: Bloom Energy signed roughly $7.65 billion of binding data-center fuel-cell contracts in about ninety days plus a $25 billion Brookfield commitment, as solid-oxide fuel cells went from backup to primary generation. The nuclear operators — Constellation, Vistra, Talen — monetize existing fleets through twenty-year PPAs now (Microsoft–Constellation, AWS–Talen, Meta–Vistra), while the SMR developers are a 2030-plus optionality play the market keeps mispricing as near-term. And liquid cooling (Vertiv, and Eaton after its $9.5 billion Boyd Thermal purchase) rides the rack-density curve. The structural second-order effect is that power has become strategic enough to pull hyperscalers into owning generation, which re-rates a swathe of "boring" electrical-industrial names into AI-adjacent growth.

4.5 Capital: the bubble question, quantified

The capital layer is where the whole thesis is stress-tested, and the useful move is to reduce the "is it a bubble" debate to the specific numbers that will resolve it. On the spending side, JPMorgan estimates $5.5 trillion of AI capex through 2030, of which $4.1 trillion is debt-financed; Morgan Stanley sees a $1.5 trillion financing gap. On the revenue side, Sequoia's David Cahn calculates that 2026's ~$1.5 trillion of infrastructure spend needs about $3 trillion of revenue to justify, and Bain estimates the industry needs ~$2 trillion of annual AI revenue by 2030 and will likely fall ~$800 billion short. Those two sides are the bubble question, quantified.

The single number that decides it is realized GPU economic life. Michael Burry argues hyperscalers understate depreciation by roughly $176 billion across 2026–28 by depreciating chips over five-to-six years when their economic life is two-to-three; Goldman's own sensitivity shows that cutting useful life from five years to three swings cumulative 2026–31 depreciation from about $3 trillion to about $4 trillion — a $1 trillion earnings hit. If inference revenue (Goldman's 24×) and enterprise ROI inflect up before that depreciation reality bites, the capex is validated; if not, the earnings reckoning arrives. The reasoning also identifies the likely trigger: with S&P having cut Oracle to one notch above junk and its credit-default swaps at multi-year highs, the repricing is more likely to start in the credit market than the equity market — so the actionable signal is single-name AI-debt CDS, not equity multiples.

Two opportunities follow. The first is that debt is the new marginal buyer, so the private-credit originators — Apollo, Blue Owl, Blackstone — capture the yield premium of financing the build (with the tail risk that opacity hides losses until they surface). The second is the rotation: with the Magnificent Seven at ~34% of the S&P and already lagging the other 493 in the first half of 2026 (the Russell 2000 Growth index up 17%), capital is visibly de-concentrating even as spending accelerates — the AI trade is broadening beyond the mega-caps, which is itself the actionable market signal.

4.6 Geopolitics: bifurcation, the sovereign demand pool, and the Taiwan clock

The policy trajectory has settled into an unstable pattern: the US oscillates between denial and monetization (the diffusion rule rescinded, a 15% China-sales tax imposed, H200 sales nominally allowed but with essentially zero China data-center revenue actually flowing), while China builds a parallel stack. The forward evidence suggests the export-control debate may already be losing relevance: domestic accelerators reached 41% of China's market in 2025, and Morgan Stanley projects 86% by 2030. The reasoning is that restriction accelerates the very self-sufficiency it aims to prevent — Nvidia's own warning that continued curbs "hand China's market to Huawei permanently" is being borne out.

Two opportunities emerge, plus one clock. The opportunity on the American side is sovereign AI as a structural new demand pool — Saudi Arabia's HUMAIN at over $100 billion and 2,200 MW, the UAE's 5 GW Stargate, and $66 billion of sovereign AI-infrastructure deployment in 2025 — which offsets Nvidia's lost China revenue and, because it runs on CUDA, entrenches the American stack (even $100 billion Gulf budgets "can't buy their way out of Nvidia"). The opportunity on the Chinese side is the closed parallel supply chain — Huawei, SMIC, CXMT — a self-contained, capacity-constrained universe insulated from US names, whose bottleneck is SMIC yields and domestic HBM. The clock is Taiwan: with ~70% of sub-2nm capacity staying in Taiwan through 2030 despite the Arizona build-out, the single-point-of-failure does not resolve this decade, and the more meaningful de-risking is CoWoS packaging localization, not wafer fabs. The second-order effect is that the whole regime — a 15% sales tax, equity stakes, case-by-case licensing — turns chips into an instrument of statecraft, which rewards diversified, less-US-exposed toolmakers and packagers.

4.7 Historical base rates: how infrastructure booms resolve

AI is technologically new; capital cycles are not. Four historical patterns provide useful—not mechanical—base rates.

The common lesson is that useful infrastructure can coexist with poor security returns. Demand growth does not protect a supplier whose capacity, leverage or valuation assumed even faster growth. The relevant discipline is to estimate each shortage's scarcity half-life and compare it with announced capacity, qualification time and the multiple already paid.

4.8 The scenarios

The reasoning above resolves into three scenarios and a tail. The horizon is the two years ending 25 July 2028. Probabilities are analytical weights, not measured frequencies.

Probability updates should be rule-based. One company announcement is insufficient. A five-percentage-point shift requires either two independent indicators or one audited system-level change, such as a sustained hyperscaler capex reduction or a broad production-ROI dataset.

The base case (most likely) is that monetization broadens slowly. Enterprise ROI improves where workflows are deep and stays patchy elsewhere, capex stays high but decelerates, the truces hold, and no financial node breaks. Leadership rotates toward the bottleneck owners — power, memory, packaging, the neutral chokepoints, the credit financiers — while the crowded mega-caps and high-beta names trade volatile. The portfolio is the picks-and-shovels barbell, faded concentration, and quality-of-earnings screens.

The bull case is that agents cross into production. If reliability crosses the threshold and AI captures a share of labor budgets rather than software budgets, the addressable market expands an order of magnitude, MIT's 95% inverts, and demand pulls forward. Everything downstream re-rates, the application layer most of all. The tell is enterprise production deployments with audited P&L and frontier-lab gross margins turning up.

The bear case is financial, not technological. A hyperscaler cuts capex, or credit stress hits the GPU-backed debt, or an efficiency shock lands, or a lab funding round stumbles — and it transmits through the circular loop faster than a normal cycle. It does not require the technology to fail, only the money to get ahead of the earnings. The neoclouds, pre-revenue thematics, and peak-priced electrical orders fall first; the diversified hyperscalers and clean-earnings arms dealers hold up.³

The tail is Taiwan — low-probability, catastrophic, un-hedgeable, and the reason to hold any AI exposure with humility.

4.9 What the consensus may be getting wrong

Four places where the crowd looks mispriced. It may be under-pricing the power constraint's durability — the physics of transformers and interconnection queues suggests it persists for years, which keeps the electrical complex a longer trade than consensus assumes. It may be over-pricing the pre-revenue nuclear and humanoid stories, which are 2030s payoffs trading as if they solve the 2027 crunch. It may be under-appreciating how far the developer and open-weight layer has already gone Chinese, which undercuts the "US leads AI" framing and matters for the third-bloc contest. And it is almost certainly under-weighting the credit vector — the shift from equity to debt is the least-discussed and most systemic feature of the build, and the regime change is likelier to show up in spreads than in stock prices.

4.10 The one number, and the calendar

If an investor tracks one thing, it should be whether disclosed hyperscaler AI revenue begins to cover the depreciation on what has been built — the metric that resolves the bubble question by settling the realized-GPU-life debate. Around it, the near-term calendar is unusually dense with resolving events.

The next several quarters will resolve most of the open questions of this atlas — the durability of capex, the trajectory of controls, the pace of the bottleneck easing. An investor who watches these signposts, and above all the depreciation-versus-revenue number and the credit spreads, will see the regime change in either direction before the market fully prices it.

4.11 Monitoring dashboard

No single disclosed metric fully measures AI's economic return. The operating dashboard therefore uses a linked set of indicators.

Every actionable security in Part V maps to at least two of these indicators and a stale-after date. This makes the atlas refreshable without rewriting its entire narrative whenever a quarterly number changes.

Sources

Linked evidence for this chapter's figures and load-bearing claims: ^{3 1 2}