AI's Three Bottlenecks: Chips Now, Power in 2028, and a $1.4T Capital Gap
J.P. Morgan's new AI infrastructure report carries its thesis in the title: "Infinite Tokens, Finite Atoms." Token demand keeps compounding, but the physical inputs that produce tokens — wafers, power and capital — do not. The report sizes each constraint, and the numbers are worth walking through.
KEY TAKEAWAYS
1. Google now processes 3,200 trillion tokens a month across its surfaces — roughly 330x the May 2024 level. Growth accelerated, not slowed, as agentic AI arrived.
2. The binding constraint today is chips, not power: TSMC N3/N2 wafer capacity and CoWoS packaging. J.P. Morgan does not see supply-demand balance before 2028.
3. The buildout needs roughly $5.5 trillion through 2030. Operating cash flow and existing channels cover about $4.1 trillion, leaving a $1.4 trillion gap for alternative capital.
Token demand is still accelerating
Start with the demand side, because it settles the "is AI demand real" question with a public number. At Google I/O 2026 on May 20, Sundar Pichai said Google's surfaces processed 9.7 trillion tokens a month in May 2024, roughly 480 trillion in May 2025, and over 3,200 trillion in May 2026. That is about 330x in two years.
J.P. Morgan splits the curve into three phases — chatbots (~5x), reasoning models (~4x), agentic AI (~10x) — and notes each phase grew faster than the last. The workload mix is tilting toward inference, which the bank estimates at 60% of workloads in 2026. Enterprise adoption is early but firm: its survey has AI at 4.5% of IT budgets today, heading to 5.8% over the next 12 months, with 56% of US firms paying for AI subscriptions (81% in tech).
Constraint 1: Compute — wafers first, packaging next
The first bottleneck is leading-edge wafer capacity at TSMC. On the bank's estimates, N3 output rises from 763K 12-inch-equivalent wafers in 2024 to 3,192K in 2028, and N2 from 316K in 2026 to 2,086K in 2028. Even at that pace, J.P. Morgan does not expect supply and demand to balance before 2028.
The company-level data points in the same direction. At its Q2 2026 earnings call, TSMC lifted full-year capex guidance to $60–64 billion from $52–56 billion, said 3nm lines were running above 100% utilization, and named CoWoS as the primary bottleneck. TrendForce-cited estimates have the CoWoS supply gap narrowing from about 20% to about 10% by end-2026 — narrowing, not closing.
Memory: HBM gets relief, commodity DRAM gets the demand
Memory sits on the constraint list too, but with a twist. HBM remains short, yet two trends are easing the pressure: the "de-spec" pattern of trimming HBM content per accelerator, and KV cache offloading to NAND. Nvidia's CMX platform, announced at CES in January 2026, standardized the tiering of inference context from GPU HBM down through CPU DRAM to NVMe SSD.
The more interesting call is on general-purpose DRAM. J.P. Morgan expects agentic AI to drive commodity DRAM demand hard through 2027: total DRAM bit growth of 29% in 2026 and 32% in 2027, while HBM bit growth decelerates from 76% in 2026 to 44% in 2028. Agentic CPU demand stays largely unmet in 2026, with CPU units compounding at 38% a year from 2025 to 2028 on the bank's numbers.
Constraint 2: Power — a 2028-29 problem, not a 2026 one
Power gets the headlines, but J.P. Morgan's read is that it is not the binding constraint yet. US installed data center capacity is projected to double by 2030 and quadruple by 2035, with forecasts still being revised up. Meanwhile the workaround menu keeps growing: behind-the-meter generation, bitcoin mine conversions, non-US sites, even orbital data centers.
One number stands out: the share of data centers planning 100% on-site generation jumped from 1% to 33%. If the grid cannot deliver, operators will generate their own. On that basis, the bank pegs the handoff of the bottleneck from chips to power at around 2028-29.
Constraint 3: Capital — a $1.4 trillion gap
The last constraint is money. J.P. Morgan sees hyperscaler capex more than doubling in 2026 and potentially reaching $3 trillion a year by 2030 at the top end, and argues the market has systematically under-forecast capex for years. The tracked guidance agrees on direction: the big four (Alphabet, Microsoft, Amazon, Meta) are guiding to roughly $725 billion combined for 2026, up 77% from about $410 billion in 2025.
Total funding needs run to about $5.5 trillion through 2030. Operating cash flow and existing channels cover roughly $4.1 trillion, leaving about $1.4 trillion for alternative capital — private credit, securitization and the like. Most hyperscalers turn free-cash-flow negative from 2027, but the bank sees the balance sheets as manageable: net debt at 13% of capital, roughly $1.7 trillion of remaining investment-grade issuance headroom, and equity issuance already up sharply to $180 billion in 2026.
The unit economics have improved too. Building 1GW of AI infrastructure costs about $42 billion (including $28 billion of IT hardware). Model and token sellers can generate roughly $31 billion a year per GW at a 44% gross margin — up from a $10 billion estimate a year earlier — while public clouds selling GPU compute earn about $17 billion at 54%. GPU rental spot prices trading above legacy contract rates support the revision.
What I actually watch
| Checkpoint | Why it matters |
|---|---|
| TSMC capex and CoWoS commentary each quarter | Whether balance arrives before 2028 sets the length of this cycle |
| Commodity DRAM bit demand | Confirms or refutes the agentic-AI-drives-DRAM call; challenges the HBM-only frame |
| Hyperscaler FCF and funding mix | How far private credit and securitization rise after the 2027 FCF flip |
| On-site generation share and BTM deals | Leading indicator for the 2028-29 chip-to-power bottleneck handoff |
Value chain read-through
| Segment | Read-through |
|---|---|
| Leading-edge foundry and packaging | Supplier's market through 2028 on the report's timeline |
| Memory | HBM growth decelerates; commodity DRAM demand stays firm through 2027 |
| Power equipment and generation | The bottleneck handoff benefits arrive around 2028-29, not now |
| Capital markets | $1.4 trillion of alternative capital demand — watch bond and securitization issuance |
Risks to this view
- Most figures above are J.P. Morgan estimates; capex and bit growth forecasts get revised every quarter.
- Token demand and revenue are different things. Whether recursive self-improvement accelerates or dampens the investment cycle is a question the report leaves open.
- A data center or power crunch could translate into AI accelerator inventory building.
- If HBM de-spec and NAND offloading move faster than expected, HBM supply could loosen ahead of schedule.
Next up: the report's fifth debate — what growing open-source token share means for the capex cycle.
Sources: J.P. Morgan, "Infinite Tokens, Finite Atoms"; Google I/O 2026 keynote (May 20, 2026); TSMC Q2 2026 earnings call (July 16, 2026); TrendForce (June 15, 2026); Nvidia CMX announcement (CES, January 2026).
Comments
Post a Comment