AI Infrastructure: Where the Bottleneck Moves Next

The Argument in Brief

AI compute has been limited by a sequence of scarce inputs: first the accelerator itself, then the advanced packaging that joins logic to memory, then the memory, and now the land, grid connection and electrical gear that let a finished rack run. Whoever holds the tightest input earns the widest margin, and the margin moves as the constraint does.

Third-quarter 2026 filings and earnings releases support that, with one correction. The constraints did not hand off one at a time. They stacked. Packaging is easing but still tight, memory has become the most profitable choke point in the chain, and power has the longest lead time and the least capacity to be solved by spending. Over the next two years, the layer that can add capacity fastest loses its pricing power first.

Demand is not in doubt this quarter. Four hyperscalers reported combined capital spending of about $203 billion in a single quarter (sources in the next section), and Alphabet and Amazon both said they expect to spend more in 2027. The risks sit in how that spending is financed and in how long suppliers’ order books stay ahead of what customers can actually energize.

Hyperscaler Capex Sets the Pace

Quarterly capital spending at the four largest buyers is now larger than the annual spending of most industrial sectors. Amazon reported $77.3 billion in the quarter, up from $35.1 billion a year earlier. Alphabet reported $51.36 billion, below the analyst estimate of $55.09 billion. Microsoft’s cash additions to property and equipment were $39.4 billion in its fiscal first quarter (ended September 30, 2026), double the $19.4 billion of a year earlier. Meta reported $34.91 billion including principal payments on finance leases.

Full-year guidance points the same way. Alphabet kept its 2026 forecast in a range that CNBC reports as $195 billion to $205 billion, and its finance chief said capex should “increase significantly in 2027.” Amazon held about $220 billion for 2026, and Andy Jassy said Amazon would “spend meaningfully more in 2027.” Meta narrowed its 2026 range to $135 billion to $145 billion. Those three alone imply $550 billion to $570 billion for the year. We leave Microsoft out of that sum because its fiscal year ends in June and we could not tie a calendar-year figure to a filing.

Customers are buying capacity ahead of revenue, and the revenue is arriving. Microsoft’s commercial remaining performance obligation (contracted revenue not yet recognized) was $677 billion, up 73%, and Azure grew 48%. AWS sales rose 39% to $45.98 billion. Google Cloud revenue more than doubled to $30.7 billion. Demand for compute is growing faster than the capital that serves it, which is the condition under which a bottleneck forms.

Two details matter for what follows. First, Microsoft disclosed leases that have not yet started of $364.6 billion, mostly for data centers, commencing between fiscal 2027 and fiscal 2034. Buyers are reserving physical capacity years out, which is what a scarce input looks like from the demand side. Second, the money is no longer free. Meta’s operating cash flow was $32.79 billion against its $34.91 billion of capex, leaving free cash flow of negative $2.12 billion. We return to that in the section on what ends the cycle.

The Accelerator Is No Longer the Scarce Part

Accelerator supply defined the first phase. NVIDIA’s second quarter of fiscal 2027 (ended July 26, 2026) produced revenue of $96.2 billion, up 106%, with Data Center at $89.0 billion. Its third-quarter outlook is $108.0 billion, plus or minus 2%, with no China data center compute assumed. NVIDIA’s Vera Rubin generation began production shipments in the third quarter of fiscal 2027, and the company says it is “currently experiencing certain supply constraints.”

NVIDIA still prices like a scarcity holder. Gross margin was 75.0% in the quarter, against 72.5% a year earlier, and operating income was $63.7 billion on those revenues, a margin of 66.2%. A year before, the same ratio was 60.8% ($28.4 billion on $46.7 billion). The margin rose while the supplier base for its inputs grew more demanding, so NVIDIA is passing cost through.

The more informative change is in who NVIDIA sells to. Hyperscale customers contributed $48.7 billion, up 102% year over year, which is about 55% of Data Center revenue. The rest came from AI clouds, enterprises, industrial and sovereign buyers, a group that grew 138% to $40.3 billion. NVIDIA’s large customers are the ones also building their own chips, so the growth in the second group shows where the merchant market is widening.

The scarce part of the system is no longer the GPU die. The same filing that reports record revenue spends several paragraphs on what lies outside the chip, which brings us to the memory, packaging and power layers.

Custom Silicon Takes a Growing Share

The largest buyers have been designing their own accelerators with partners, and Broadcom is the visible beneficiary. In its third fiscal quarter (ended August 2, 2026), Broadcom’s AI semiconductor revenue was $16.7 billion, up 221% from a year earlier and up 54% from the prior quarter. Management guided the fourth quarter to $21.7 billion of AI semiconductor revenue. That guide is about 62% of the $34.8 billion company revenue forecast, so Broadcom is now mostly an AI company by revenue.

Scale keeps the comparison honest. NVIDIA’s Data Center line was more than five times Broadcom’s AI line in the latest quarters, and the two report on different calendars. Custom silicon is growing faster in percentage terms from a much smaller base, and it does not replace NVIDIA in the near term. It changes the negotiation. A buyer with a credible in-house alternative has bargaining strength on price, and the first evidence of that would be NVIDIA’s gross margin falling below the 74% it guided for next quarter.

Broadcom’s own margin shows the same ordering as NVIDIA’s, one step down. Its GAAP operating income was $16.0 billion on $29.6 billion, a margin of 53.9% versus 36.9% a year earlier. Custom silicon earns a lower margin than merchant GPUs because the customer holds more of the design value, yet it is expanding rapidly.

Foundry and Packaging: Easing, Still Tight

TSMC makes a large share of the leading AI chips, merchant and custom. In the third quarter of 2026, TSMC’s revenue was US$46.72 billion, up 41.2% from a year earlier, with a gross margin of 66.4% and an operating margin of 58.7%. A year before, operating income was NT$500.7 billion on NT$989.9 billion of sales, a margin of 50.6%. Two-nanometer wafers were 8% of wafer revenue, and fourth-quarter revenue is guided to US$52.3 billion to US$54.3 billion.

TSMC reiterated its 2026 capital budget at $60 billion to $64 billion on its October 15 call and said high performance computing was 68% of revenue. The call is the best evidence we have on packaging. Chairman C.C. Wei said CoWoS (chip-on-wafer-on-substrate, the technique that mounts logic and memory side by side) is still the first choice of customers. He also acknowledged “a little bit shortage on the back end,” and said TSMC is working with GlobalFoundries to add support. We read that as a constraint that is being worked down by outside capacity, which is what a bottleneck looks like when it is moving on. We relied on a machine-generated transcript for these quotes and have not been able to confirm TSMC’s CoWoS wafer-per-month figures, so we do not quote any.

On the same call Wei said power supply is one of the factors TSMC weighs in its demand planning, and that its teams had checked U.S. data center build-outs. A foundry looking at grid capacity when it sizes its own demand is a sign of where the next limit lies.

Foundry pricing power is real, and it is already cashed in. TSMC’s margin gain of eight points in a year came with a capital budget of $60 billion to $64 billion that will relieve its own constraints over time. Pricing power at the packaging step is the most likely to fade, because the supplier is spending heavily to expand it.

Memory Became the Profit Pool

Our starting thesis listed memory as one stage among several. The latest data makes it the centerpiece. SK hynix reported preliminary third-quarter 2026 revenue of 104.6 trillion won, up 328%, and operating profit of 83.33 trillion won, up 632%. That is an operating margin near 80%, higher than NVIDIA’s, TSMC’s or any other layer in the chart below. The company said its high bandwidth memory (HBM, the stacked DRAM that sits next to an accelerator) and other DRAM are sold out through 2029, that more than half of its 2027 and 2028 capacity is covered by long-term agreements, and that it expects supply to fall short of demand through 2030. Those are management claims and a seller has reasons to describe tightness. They still line up with what NVIDIA disclosed.

NVIDIA’s supply and capacity commitments rose from $119 billion to $279 billion in a single quarter, “primarily related to the procurement of memory.” The schedule shows $92 billion due in the remainder of fiscal 2027, $87 billion in fiscal 2028 and $88 billion in fiscal 2029. The dominant accelerator vendor is committing three years of purchases to secure DRAM. A buyer with NVIDIA’s pricing power does that only when the seller has more.

The side effects are already visible outside the data center. NVIDIA said its Edge Computing revenue growth was partly offset by slower consumer PC sales, which it attributed in part to higher memory and systems prices. Memory scarcity is raising costs across the electronics supply chain, with AI buyers outbidding the consumer market.

Read the chart as a ranking of who can say no to a customer. Memory is first, accelerators second, foundry third. Each of those layers requires a new cleanroom or fab to expand, which takes years and capital the supplier has to commit before it knows demand. The bottom bar is a different type of business. Vertiv builds equipment that can be duplicated by competitors and expanded within quarters, so its margin is lower even in a boom. We come back to that contrast in the last section.

Networking: A Cost of Doing Business

Networking carried the previous bottleneck narrative. Training clusters need fast links between thousands of accelerators, and for a period the switches, optics and cables were gating deployments. The third-quarter filings give us less evidence that this holds now. NVIDIA reports its networking within a combined Compute and Networking segment, which was $88.3 billion of the $96.2 billion total, and no longer isolates it. Broadcom’s AI semiconductor figure combines custom accelerators with networking, and the company described demand for both as very strong without splitting them.

Our reading is that networking has become a proportional cost of every cluster, bought with the chips, and that it grows as fast as the build-out without commanding the distinct margin premium that memory does. We may be wrong because the disclosure is thin. If a company reports a networking line with margins clearly above its compute line, that would argue for treating networking as a separate scarcity. We could not source such a line this quarter and did not use secondary estimates.

Power and Cooling: The Slowest Constraint to Fix

The grid is where the argument becomes physical. A chip can be made in a quarter. A transformer, a turbine or an interconnection takes years.

In PJM, the largest U.S. grid operator, the capacity market has stopped clearing the reliability requirement. The 2027/28 auction in December 2025 cleared at its cap of $333.44 per megawatt-day and fell 6,623 MW short of the requirement, with nearly 5,100 MW of the 5,250 MW load forecast increase attributed to data centers. The 2028/29 auction on July 14, 2026 cleared at the $325 cap and fell 6,831 MW short, with only 525 MW of new generation and uprates offered. The market is saying that it wants more supply than anyone has built.

The longer-term demand forecasts have been revised upward. The Lawrence Berkeley National Laboratory estimated that U.S. data centers used 176 TWh in 2023, or 4.4% of national electricity, and projected a 2028 share of 6.7% to 12%. That report was published in December 2024, before most of the capex described above was announced, so we treat it as a floor on the range of outcomes rather than a forecast.

Equipment is as constrained as generation. Lead times for the largest transformers have stretched to as long as four years, according to PwC analysts quoted in May 2026, and Wood Mackenzie data show demand for generator step-up transformers up 274% between 2019 and 2025. Gas turbines show the pricing consequence most clearly. GE Vernova’s gas equipment backlog and slot reservations rose from 116 GW to 122 GW in the third quarter, and the company expects at least 130 GW by year-end. Its Power segment’s EBITDA margin was 21.4%, up 800 basis points from a year earlier, driven by “favorable price and higher volume at Gas Power.”

NVIDIA now treats energy as part of its own business. Its CFO commentary says that securing land, power and shell for data centers has become the next phase of the buildout, and that many AI clouds are growing faster than their balance sheets can support. NVIDIA has guaranteed up to $105 billion of land, power and shell obligations for roughly 4.25 GW at an Ohio campus that will host its systems under 20-year leases to OpenAI. The supplier that sold the most chips is now using its own balance sheet to buy the one input it cannot make.

Cooling and electrical equipment sit between the grid and the rack. Vertiv’s third-quarter net sales were $4,313 million, up 61% with 55% organic growth, and organic orders grew more than 100%. It raised its 2026 sales guidance by $650 million to a midpoint of $14,650 million. In February the guidance had been $13,250 million to $13,750 million, and year-end 2025 backlog was $15.0 billion with a book-to-bill of 2.9 times. Adjusted operating margin of 25.2% was up 290 basis points. Equipment makers in this layer are seeing volume and some price, and for now they are not collecting scarcity rents on the memory scale. Their moat is shorter lead times and service, which is a thinner moat than a fab.

Where Pricing Power Sits

The third-quarter data gives an ordering. Pricing power is greatest where a layer has few suppliers, long lead times to add capacity, and a buyer with no ready substitute. That describes memory first, because the DRAM makers cannot add cleanroom capacity quickly. It describes accelerators next, because NVIDIA’s software and system integration keep switching costs high even as custom silicon grows. Foundry follows, with TSMC’s lead in advanced nodes and packaging offset by its own heavy spending to relieve the constraint.

Power infrastructure is the layer most likely to gain pricing power over the next several years, and the evidence is already in the gas turbine and capacity market data. It differs from the chip layers because a hyperscaler cannot buy its way around it with a bigger check. Interconnection queues and permitting are set by regulators and utilities, and transformer capacity is set by a handful of factories. That is why NVIDIA, a company with plenty of cash, is signing guarantees on land and power.

Our original sequence of GPUs, then packaging and HBM, then networking, then power, needs two edits. The constraints overlap, so the question is which one binds in a given quarter. And the prize for relieving a bottleneck goes to the supplier that sells the scarce item, rather than the one that builds the equipment around it. Vertiv’s margin illustrates the second edit. It is growing as fast as any company in this report and earns less than half of NVIDIA’s operating margin.

What Would End the Cycle

We list the signals that would change this view. None has occurred, and each is measurable.

Financing runs ahead of cash generation. Meta’s free cash flow was negative in the third quarter. Amazon’s capex of $77.3 billion in one quarter is more than double its year-earlier level. When spending exceeds internally generated cash at the largest buyers for several quarters, they need external capital, and the cycle becomes sensitive to credit markets. The next check is the fourth-quarter reports due in early 2027, which will show whether free cash flow turns positive at the four companies in the first chart.

Circular financing at the suppliers. NVIDIA’s commitments total $366 billion, including $25 billion of equity investments in model makers and infrastructure financiers, and its guarantees on customer leases cap at $108.5 billion. When a vendor finances its customers’ purchases, reported demand overstates independent demand. A rise in receivables days beyond the 60 days NVIDIA reported would be an early sign.

Customer revenue stops covering the commitments. Microsoft’s $364.6 billion of future lease commitments are fixed obligations. If cloud growth slows from 48% on Azure to a rate below the growth of capex for several quarters, the commitments turn into stranded cost. This is the most direct way for the cycle to end, and the AI model makers that buy capacity are the least transparent part of the chain.

Memory supply catches up. SK hynix says shortage runs to 2030, which is a long way off and a statement from the seller. If contract memory prices fall for two consecutive quarters, the margin premium in the chart would compress and NVIDIA’s purchase commitments would look expensive in hindsight.

Power arrives faster than expected, or demand needs less of it. A fast, funded resolution of the 6.8 GW shortfall in PJM would lessen the scarcity. A drop in energy per unit of compute from efficiency gains would also relax it. Either would shift pricing power away from the power layer.

Custom silicon takes the margin. If NVIDIA’s gross margin falls below 70% while Broadcom’s AI revenue keeps growing, buyers are winning the negotiation. NVIDIA guided to 74% next quarter.

The Companies

NVIDIA (NVDA)

NVIDIA sells the accelerators, systems and networking that anchor most AI clusters, and it reports record revenue of $96.2 billion for the quarter ended July 26, 2026. Its position in this report’s argument is two-sided. It captures a 66% operating margin from the accelerator layer, and it is now a large buyer of memory, with $279 billion of supply and capacity commitments. It has also begun to finance the downstream constraint through lease guarantees. The watch items are gross margin against the 74% guide and the share of revenue from customers with in-house chips. Read the full NVIDIA memo.

TSMC (TSM)

TSMC manufactures and packages a large share of advanced AI silicon, with third-quarter revenue of US$46.72 billion and 68% from high performance computing. It is the layer where the bottleneck has been easing most visibly, with a capital budget of $60 billion to $64 billion and outside packaging support. Its margin is high, and its main risk to the thesis is that capacity expansion works and ends the premium. See the Taiwan Semiconductor memo.

Broadcom (AVGO)

Broadcom designs custom AI accelerators and networking chips for large buyers, and AI semiconductors reached $16.7 billion in its fiscal third quarter, with a $21.7 billion guide for the fourth. It is the clearest read on whether buyers are shifting spend toward in-house designs. Customer concentration is the offsetting risk, and the fiscal third-quarter 10-Q is the place to check it. See the Broadcom memo.

Vertiv (VRT)

Vertiv supplies power distribution, backup power and cooling equipment for data centers. It grew third-quarter sales 61% and more than doubled organic orders, which makes it the cleanest listed expression of the shift in the bottleneck toward the building. Its margin sits well below the chip layers, so its earnings depend on volume and execution more than on scarcity pricing. See the Vertiv memo.

GE Vernova (GEV)

GE Vernova makes the gas turbines and grid equipment that determine how fast new load can be served. Its gas backlog and slot reservations stand at 122 GW, and its Power segment margin rose 800 basis points on price. It is the best evidence in the filings that power generation has started to earn scarcity rents. See the GE Vernova memo.

SK hynix

SK hynix is a Korean-listed memory maker that we do not cover with a memo. Its third-quarter operating profit of 83.33 trillion won on 104.6 trillion won of revenue is the strongest single data point for where the profit pool sits today. The statement that supply is sold out through 2029 is management’s and should be checked against contract pricing as it reports.