The AI Data Center Gold Rush: Who's Selling the Shovels?

maktinta
Jun 25
17 min read

A fact-checked and expanded field guide to the AI infrastructure supply chain

Figures verified and refreshed as of June 2026. Sources listed at the end.

Every gold rush in history has the same cast of characters: the lucky few who strike it rich, and the many who get rich selling shovels, pickaxes, and denim pants to everyone headed to the mine.

The AI infrastructure boom is no different. But there is one important update to the old story. In the original gold rush there was only one kind of gold. In this one, the gold itself now comes in several forms and that is where we have to start.

The Gold : It's No Longer Just NVIDIA

NVIDIA is still the gold standard, and it is still winning. In its most recent quarter (Q1 FY2027, reported May 2026), NVIDIA posted record data center revenue of $75.2 billion, up 92% year-over-year, on total revenue of $81.6 billion.

It holds somewhere between roughly 80% and 90% of the merchant AI accelerator market depending on how you count, and well above 90% of discrete GPUs. The H100 and H200 are now the previous generation; the Blackwell and Blackwell Ultra (GB300) racks are what hyperscalers are fighting over today. Lead times have eased from the worst of 2024, but the newest rack-scale systems are still effectively sold out and allocated quarters in advance.

What has genuinely changed since this story was first told is that NVIDIA is no longer the only credible source of AI compute. A field of established alternatives is now shipping at real scale — some merchant (you can buy them), some captive (a hyperscaler builds them for itself). Anyone mapping this supply chain needs to know the whole field, not just the leader.

The challengers, briefly

AMD is the clear number-two merchant GPU vendor. Its Data Center segment hit $5.8 billion in Q1 2026, up 57% year-over-year, carried by the Instinct MI350 ramp. At CES 2026 AMD detailed the MI400 series (MI430X / MI440X / MI455X) and the Helios rack-scale system — 72 MI455X accelerators, ~31 TB of HBM4, and roughly 3 AI exaflops per rack — its first true answer to NVIDIA's NVL rack. Helios and MI400 production are slated for the second half of 2026, with more detail due at AMD's July "Advancing AI" event. AMD's AI-accelerator share is still in the mid-single digits (~5–7%), but analysts at S&P Global Market Intelligence project around $7.2 billion of MI400-series revenue alone.

Google TPU is the most important non-NVIDIA accelerator on the planet. Google made its seventh-generation TPU, Ironwood (TPU v7), generally available in April 2026 and positioned it as "inference-first." The headline proof point: Anthropic committed to up to one million TPU chips and more than a gigawatt of capacity in 2026, scaling toward an additional 3.5 GW in 2027 — one of the largest compute deals ever signed. Notably, the TPU is co-designed with Broadcom (see Section 1), so a TPU win is also a Broadcom win.

AWS Trainium is Amazon's captive answer. Trainium2 reached general availability in late 2024, and Trainium3 follows with materially higher throughput and memory bandwidth. Amazon's "Project Rainier" build-out for Anthropic — backed by Amazon's $8 billion investment in the company — is one of the largest custom-silicon clusters in existence.

Cerebras takes the opposite architectural bet: instead of thousands of small chips, a single dinner-plate-sized wafer. Its CS-3 system, built on the WSE-3 wafer-scale engine, anchors an inference business that reached $510 million in 2025 revenue (up 76%) and was actually profitable — rare for an AI-chip startup. Cerebras landed a 750 MW capacity agreement with OpenAI and went public in May 2026, raising about $5.5 billion at a ~$23–27 billion valuation with the stock roughly doubling on debut. Its next-generation wafer-scale part is on the roadmap.

Groq specializes in ultra-low-latency inference with its LPU (Language Processing Unit). It raised $750 million in September 2025 at a $6.9 billion valuation and sells both GroqCloud and on-prem LPX racks; its newest LPU targets agentic workloads at very high token-per-second rates.

TensorWave is the neocloud built specifically on AMD. In June 2026 it raised a $350 million Series B at a $1.55 billion valuation (co-led by AMD Ventures), has deployed 8,192 MI325X GPUs, and has secured ~2 GW of long-term data-center capacity for MI355X clusters. It is, in effect, the AMD-flavored answer to the NVIDIA-centric neoclouds.

And the rest of the bench: Intel's Gaudi 3 (notable for 24 integrated 200 GbE ports per chip, which helps push the whole industry toward Ethernet), Meta's in-house MTIA, Microsoft's Maia, plus Huawei Ascend in China and a long tail of startups (SambaNova, d-Matrix, Tenstorrent). The strategic point: every one of these still needs the same shovels — memory, optics, fiber, switching, power, cooling, and land. The compute layer is diversifying; the picks-and-shovels layer underneath it is converging.

The AI compute field at a glance

Player	Flagship / latest	Scale & growth	Status
NVIDIA	Blackwell Ultra (GB300) racks	$75.2B data-center rev, +92% YoY (Q1 FY27); ~80–90% accel. share	Shipping; high-end still allocated
AMD	MI350 now; MI400 + Helios rack	$5.8B data-center rev, +57% YoY (Q1 26); ~5–7% share	MI400/Helios ramp H2 2026
Google TPU	Ironwood (TPU v7)	Anthropic: up to 1M chips, >1 GW in 2026	GA April 2026; co-designed w/ Broadcom
AWS Trainium	Trainium2 GA; Trainium3	Anchors Project Rainier for Anthropic	Captive; scaling fast
Cerebras	CS-3 (WSE-3 wafer-scale)	$510M 2025 rev, +76%; profitable	IPO May 2026 (~$5.5B raised)
Groq	LPU / LPX (inference)	$6.9B valuation (Sep 2025)	GroqCloud + on-prem racks
TensorWave	AMD MI355X neocloud	$1.55B valuation; ~2 GW capacity	$350M Series B, Jun 2026

Bottom line: NVIDIA is the gold, but the gold now comes in several alloys. The rest of this guide is about who sells the shovels every one of them depends on.

The Server Builders

A loose GPU is just an expensive paperweight. Before an accelerator can do any work it has to be bolted into a server, wired to memory and networking, plumbed for liquid cooling, tested, and shipped as a working system. That assembly job has quietly become one of the biggest and fastest-growing businesses in the entire build-out. The generative-AI server market was worth about $104 billion in 2025 and is projected to reach roughly $449 billion by 2030 — and it splits into two camps: the brand-name OEMs you have heard of, and the Taiwanese ODMs who quietly build most of the racks.

Dell is the OEM poster child.

In fiscal 2026 it booked more than $64 billion in AI-optimized server orders, shipped over $25 billion, and entered FY2027 with a record $43 billion backlog. Its Infrastructure Solutions Group posted record full-year revenue of $60.8 billion (up 40%), with AI-optimized server revenue reaching $9.0 billion in a single quarter. Dell’s edge is scale, financing, and enterprise reach — it can deliver an NVIDIA-based cluster as a supported, warrantied system rather than a pile of parts.

Super Micro (SMCI) is the specialist that got there early. It guided fiscal 2026 revenue to $38.9—40.4 billion (around 65% growth) and runs what it calls the world’s largest liquid-cooled rack production — around 5,000 racks per month. More than 70% of its revenue now comes from GPU/AI platforms, and its Blackwell order book ran near $13 billion early in 2026. Supermicro’s bet on direct-liquid cooling — once seen as exotic — turned out to be exactly what 100 kW-plus racks require.

HPE and Lenovo round out the Western field. HPE’s AI-infrastructure revenue grew 52% year-over-year in fiscal Q2 2026 on what its CEO called the strongest AI-server backlog the company has ever seen, and its Juniper-powered networking segment grew nearly 150% — letting HPE sell the compute and the network together. Lenovo’s Infrastructure Solutions Group reached $19.2 billion in annual revenue, with AI-related revenue up 84% year-over-year and now more than a third of the whole company.

But the companies that actually screw most of these racks together sit in Taiwan, and together they are bigger than all the brands combined. Hon Hai (Foxconn) assembles roughly half of the world’s NVIDIA GB200/GB300 rack-scale systems — about 52% share in 2025, with management guiding toward 55—60% in 2026 — and has set a 2026 revenue target near $350 billion. Quanta (about 19% of GB-rack share) expects triple-digit AI-server growth, and Wistron/Wiwynn holds around 20%. These ODMs are the true mass producers of AI compute; the brand-name OEMs often design, support, and finance the very systems the ODMs build.

The strategic point mirrors the rest of this guide: whether the gold inside is an NVIDIA Blackwell, an AMD Helios, or a Google TPU, someone has to box it, cool it, cable it, and ship it as a working machine. The server builders sell that shovel to every miner at once — and right now they cannot assemble fast enough.

Building the AI Internet

If NVIDIA is the engine, Broadcom is the road the engine drives on. Broadcom makes two things essential to every large-scale AI deployment: custom AI chips (XPUs / ASICs) and the network switches that connect them.

On the custom-chip side, Broadcom co-designs application-specific accelerators with hyperscalers. Google's TPU — the chip behind Search, Translate, and Gemini, and now Anthropic's million-chip commitment — has been co-designed with Broadcom across seven generations dating to 2014. Google is one of several confirmed XPU customers Broadcom has publicly acknowledged.

The numbers tell the story. Broadcom's AI semiconductor revenue hit $10.8 billion in Q2 FY2026, up 143% year-over-year, and the company guided Q3 AI revenue to $16.0 billion — over 200% growth year-over-year. Total Q2 revenue was a record $22.2 billion. (Both figures verified against Broadcom's earnings release.)

On networking, Broadcom's Tomahawk 6 switch chip entered volume production in early 2026 as the first 102.4 terabit-per-second Ethernet chip, and its companion Jericho 4 fabric chip is designed to interconnect more than a million XPUs across data centers. That is not a data center — that is a city of silicon, and Broadcom is building the highway system. In merchant data-center switch silicon, Broadcom remains the dominant supplier.

The Memory Tax on Every AI Thought

A GPU can only process data as fast as it can access memory. That is why High Bandwidth Memory (HBM) — RAM stacked directly on the accelerator — has become one of the most strategically important components in the entire AI supply chain.

There are only three HBM makers in the world: SK Hynix (the share leader, around half the market), Samsung, and Micron. Micron sits third but is gaining, and demand is eating supply alive. In Q2 FY2026, Micron's total revenue hit a record $23.8 billion, up roughly 196% year-over-year (from about $8.05 billion a year earlier). Its Cloud Memory unit, where HBM sits, generated $7.7 billion alone, up 163%. Micron's entire 2025 HBM output was sold out before the year began, and it guided Q3 FY2026 revenue to roughly $33.5 billion.

On product: Micron began volume shipments of 36 GB, 12-high HBM4 in early calendar 2026, aligned to NVIDIA's Vera Rubin roadmap, and is already sampling a 16-high part that pushes capacity to 48 GB per cube.

Every NVIDIA GPU, every Broadcom XPU, every AMD Instinct, and every TPU ships loaded with HBM. Micron doesn't get the headlines, but it collects a toll on every AI thought that gets processed.

Light Speed Is Not a Figure of Speech

Inside an AI data center, thousands of accelerators must talk to each other constantly and incredibly fast. Past a certain distance and data rate, copper runs too hot and burns too much power. The answer is light, and optical transceivers are the nervous system of the modern AI data center.

Coherent is one of the dominant transceiver makers. In a landmark move, NVIDIA made a $2 billion equity investment in Coherent alongside a multi-billion-dollar, multi-year purchase commitment to secure capacity for advanced lasers and optical networking through the rest of the decade. That is NVIDIA buying a stake in its own circulatory system.

Coherent's Data Center & Communications segment now accounts for about 75% of total revenue and grew 41% year-over-year to roughly $1.36 billion in Q3 FY2026, on record total revenue of $1.81 billion. The roadmap runs through 800G and 1.6T transceivers and into co-packaged optics (CPO), where the optics move into the chip package itself; Coherent expects scale-out CPO revenue to begin in the second half of 2026.

Fiber Frenzy: You're Going to Need a Lot More Glass

Before data can travel at the speed of light, it needs something to travel through: fiber-optic cable, hair-thin strands of ultra-pure glass. An AI GPU rack needs on the order of 36 times more fiber than a traditional CPU rack — multiply by thousands of racks per campus and the math gets dramatic.

In January 2026, Meta signed a deal to pay Corning up to $6 billion for fiber to support its AI build-out. It was not the only one: in June 2026 Amazon signed its own multi-billion-dollar fiber agreement with Corning, and Corning disclosed additional large, long-term deals with unnamed hyperscalers. Corning's enterprise optical sales grew 58% year-over-year, with total Optical Communications up about 36%, driven almost entirely by AI demand. The company even invented a new AI-specific fiber, Contour, that doubles strand density in a standard conduit.

The broader fiber-optic cable market, valued at roughly $13.9 billion in 2025, is projected to reach about $20.9 billion by 2030. Italy's Prysmian, the world's largest cable maker, is aggressively expanding U.S. fiber capacity. At least one major manufacturer has sold its entire 2026 inventory — the glass is moving before it's even made.

The Switchboard That Holds It All Together

If fiber is the highway and transceivers are the on-ramps, network switches are traffic control. Arista has taken the #1 share position in high-speed (>10G) data-center switching, passing Cisco — a milestone for a company most people outside networking have never heard of. Full-year 2025 revenue grew 28.6% to $9.01 billion; Q1 2026 was up again (~27–35% depending on the line), and Arista expects to exceed $10 billion in revenue in 2026, with AI-specific revenue targeted around $2.75 billion.

Cisco isn't standing still: it raised its AI infrastructure revenue guidance for fiscal 2026 from $2 billion to $3 billion. The architectural shift driving both: AI clusters are moving from proprietary interconnects toward standard Ethernet fabrics. With even Intel's Gaudi 3 carrying 24 integrated 200 GbE ports per chip, the industry is converging on Ethernet — and Arista and Cisco are the beneficiaries (with NVIDIA's own Spectrum-X Ethernet now competing directly on that turf).

The Unglamorous Makes Everything Possible

Nobody tweets about power distribution units, but without them, every GPU goes dark. Vertiv is the picks-and-shovels play of the power-and-cooling world. Q1 2026 revenue hit $2.65 billion, up 30% year-over-year, and full-year guidance sits at $13.5–14.0 billion. Notably, its order backlog has more than doubled to over $15 billion — a meaningful jump from the ~$9.5 billion figure quoted in the original piece, and a sign of just how fast demand is accelerating. Vertiv is one of the only firms covering both power and cooling, which matters as rack densities push past 100 kW.

Eaton and Schneider Electric round out a trio that together controls a large share of the data-center power-equipment market. The data-center UPS market alone is projected to grow from $8.76 billion in 2025 to $12.47 billion by 2030. At gigawatt scale, a one-second interruption isn't an inconvenience — it's a catastrophic event. Bulletproof power distribution is not optional.

The Hidden Bottleneck : Power Transformers

Here is the wildcard quietly threatening the entire build-out. A large AI data center needs massive transformers to step utility transmission voltage down to usable levels. These are custom-engineered, heavy industrial equipment — and they take a very long time to build.

Before 2020, lead times for high-power transformers ran about 24–30 months. Today they have stretched to three to five years in some cases, while AI projects are promised on 12–18-month timelines. The math doesn't work. As a result, nearly half of planned U.S. data-center builds for 2026 are expected to be delayed or canceled. Roughly 12 GW of U.S. capacity is slated to come online in 2026, but only about a third is currently under active construction — not for lack of capital or chips, but for lack of transformers.

Chinese manufacturers control roughly 60% of global transformer production capacity, with Chinese transformer exports hitting a record in 2025 (up ~36% year-over-year by value). Geopolitical risk is embedded directly into the AI supply chain, in a piece of equipment most people couldn't pick out of a lineup. The companies that can secure transformers — or that control sites already connected to the grid, have a structural advantage no amount of capital can quickly replicate.

The Thermal Architects of the AI Boom

Every watt that goes into an AI data center becomes heat, and that heat has to go somewhere. The chiller and dry-cooler industry is facing a demand surge it was not built to expect.

Trane Technologies moved first, launching a thermal-management reference design built around NVIDIA's Omniverse DSX blueprint — embedding itself into the design standard for gigawatt-scale AI factories. Johnson Controls followed with a second AI Factory Reference Design Guide in May 2026, focused on air-cooled chiller architectures scalable to a 1 GW campus. Carrier introduced high-ambient-temperature chillers purpose-built for AI workloads.

And Modine — a company most people couldn't place — is quietly becoming one of the most important thermal suppliers in the industry. It commissioned four new chiller production lines in a single quarter in 2026, with four more scheduled (eight total), as data-center sales rose 31% sequentially in fiscal Q3 2026 on quarterly revenue exceeding $400 million. The physics of AI compute leave no room for afterthought cooling — and none of these firms can build capacity fast enough.

Generators, Batteries, the Show Must Go On

An AI data center going offline is a catastrophic event — models lose training state, inference drops requests, and the financial exposure can run into tens of millions of dollars per hour. The backup-power industry is having a moment unlike anything in its history.

Caterpillar and Cummins dominate industrial diesel generators, and both are now effectively AI infrastructure plays. UBS projects each could add up to $1.5 billion in new revenue from data-center generators alone. Caterpillar's power-generation retail sales surged 44% year-over-year; Cummins raised its full-year revenue forecast from 8% to 11%, driven largely by data-center demand; and Caterpillar shares rose roughly 45% in 2025 — without the company making a single GPU. The data-center generator market is valued around $8.57 billion in 2026, growing to ~$9.79 billion by 2031.

But diesel is only part of the story. As campuses reach gigawatt scale, operators layer in battery energy storage (BESS) for instantaneous backup and to smooth the spiky loads AI training creates. Tesla's Megapack is a leading solution — its Shanghai factory targets 40 GWh of annual capacity, and the new Megablock bundles four Megapack 3 units with an integrated transformer and switchgear into a 20 MWh plug-and-play skid. CATL, the world's largest battery maker, invested $600 million in 2026 in HVDC-compatible storage aimed specifically at data centers, and Fluence Energy is expanding data-center deployments through its optimization platform.

The Infrastructure Nobody Tweets About

Between the substation and the server rack sit dozens of components most people never think about: automatic transfer switches, medium-voltage switchgear, busways, variable-speed pump skids, glycol tanks, thermal buffer vessels. None of it is glamorous; all of it is indispensable; and right now virtually all of it is on backorder. ABB, Siemens, and GE Vernova supply the switchgear that routes power through the facility, and Siemens Energy partnered with Eaton in 2025 specifically to enable simultaneous construction of data centers and on-site generation — an admission that the grid alone can no longer be the only plan.

On the fluid side, Grundfos and Xylem dominate intelligent pump systems for liquid-cooled data centers — variable-speed, IoT-enabled units that modulate flow in real time across thousands of cooling loops. The data-center cooling-pump market is valued around $2.5 billion in 2025 and is projected to exceed $4 billion by 2033 (~8% annually) — a quiet, steady winner most investors haven't thought to look at. And the tank fabricators welding giant glycol buffer and thermal-storage vessels? Mostly not publicly traded, and fully booked.

Construction Companies & Real Estate Developers

At some point, someone has to build the thing. The global data-center construction market was valued around $275 billion in 2025 and is projected to reach roughly $761 billion by 2034 (about 12% CAGR; estimates vary by firm). In the U.S. alone, monthly construction starts hit a record $14 billion in July 2025, and starts through the first seven months reached $26.9 billion — nearly triple the comparable 2024 figure. Roughly 70% of that investment traces to just four companies: Amazon, Google, Microsoft, and Meta.

The firms building these campuses — Turner, DPR, Hensel Phelps, Mortenson — have become as specialized as aerospace contractors, coordinating complex MEP work against GPU-delivery schedules rather than typical commercial milestones. On the real-estate side, land near existing grid infrastructure has become one of the most valuable asset classes in North America.

Digital Realty, Equinix, and Iron Mountain are the landlords of the AI economy, while a new class of pure-play developer now hunts power-rich parcels, secures interconnection, and sells shovel-ready sites to operators who can't wait five years for a greenfield queue. In a market where grid access is the binding constraint, the right land in the right utility territory is worth more than almost anything else on the balance sheet.

Turbines, Cogen, Power Producers, the Nuclear Bet

Every AI data center needs power, and where it comes from has become one of the most consequential questions in energy. GE Vernova is the most direct turbine beneficiary: in Q1 2026 its Electrification segment booked $2.4 billion in data-center equipment orders — more than all of the prior year combined. Total company orders hit $18.3 billion (+71%), backlog swelled to about $163 billion, and its gas-turbine reservation backlog grew to 100 GW (targeting 110 GW by year-end). GE Vernova is expanding heavy-duty gas-turbine output and spending heavily to add capacity, offering everything from fast-start aeroderivatives to combined-cycle baseload and cogeneration that recycles waste heat.

Siemens Energy faces the same constraint and is effectively sold out; Bloomberg reported it is struggling to keep up with AI-driven turbine demand, putting hundreds of billions in planned gas plants at risk of delay. Wärtsilä has become a significant player in modular, containerized gas-engine cogeneration that deploys in phases as load grows.

On the independent-power side, Constellation Energy and Vistra are writing some of the decade's most consequential contracts. Constellation closed its $16.4 billion acquisition of Calpine in January 2026 (~$26.6 billion including debt), creating the largest electricity producer in the United States. Calpine has agreed to supply 760 MW of behind-the-meter power to CyrusOne data centers in Texas; Microsoft signed a deal with Constellation to restart Three Mile Island for AI compute; and Vistra signed ~2,600 MW of long-term nuclear power-purchase agreements with Meta across three PJM plants.

And then the wildcard: Small Modular Reactors. Google became the first company to sign an SMR power-purchase agreement, committing to Kairos Power's Hermes 2 reactor in Oak Ridge — 50 MW initially, scaling to 500 MW by 2030. Oklo and Meta agreed to a 1.2 GW nuclear campus in Ohio with site work beginning in 2026.

TerraPower's Natrium pairs a 345 MW sodium-cooled core with molten-salt storage that can surge output to 500 MW — designed precisely for spiky AI loads. The first commercial SMR-powered data centers are expected around 2030, but the deals being signed today will define who controls carbon-free baseload power well into the 2040s. Nuclear didn't come back. It never left. It just found a customer with very deep pockets and a very large power bill.

The Shovel-Seller Scorecard

A structured summary of each layer — market size, growth, leading players, and the binding constraint or lead time. Figures are the most recent verifiable as of June 2026.

Layer	Size / key metric	Leaders	Constraint / lead time
AI compute	NVIDIA $75.2B DC rev/qtr (+92%)	NVIDIA; AMD; Google TPU; Cerebras; Groq	High-end racks allocated quarters out
Servers	AI-server mkt $104B — $449B (2030)	Foxconn; Dell; Supermicro; Quanta; HPE	Foxconn ~52% of GB200/300 racks
Custom XPUs	Broadcom AI rev $10.8B→$16B/qtr	Broadcom (dominant)	Demand >200% YoY
HBM memory	Micron cloud-memory $7.7B/qtr (+163%)	SK Hynix; Samsung; Micron	Sold out through 2025–26
Optics	Coherent DC&C $1.36B/qtr (+41%)	Coherent; Lumentum	CPO ramps 2H 2026
Fiber	$13.9B (2025) → $20.9B (2030)	Corning; Prysmian	2026 inventory sold out
Switching	Arista FY25 $9.01B (+29%)	Arista (#1); Cisco; NVIDIA	Ethernet convergence underway
Power/cooling	Vertiv backlog >$15B	Vertiv; Eaton; Schneider	UPS mkt $8.8B→$12.5B by 2030
Transformers	China ~60% of capacity	Siemens; Hitachi; ABB; GE	3–5 yr lead times (was 2–2.5)
Cooling gear	Modine DC rev >$400M/qtr (+31% q/q)	Trane; JCI; Carrier; Modine	Adding lines; capacity-bound
Backup/BESS	Generator mkt $8.57B (2026)	Caterpillar; Cummins; Tesla; CATL	Gen sales +44% YoY
Construction	$275B (2025) → $761B (2034)	Turner; DPR; Hensel Phelps	70% from 4 hyperscalers
Power gen	GE Vernova DC orders $2.4B/qtr	GE Vernova; Siemens Energy; Wärtsilä	Turbines sold out to ~2030
Nuclear/SMR	Vistra–Meta ~2,600 MW	Constellation; Vistra; Kairos; Oklo	Commercial SMRs ~2030

In Summary

The AI infrastructure boom has created a supply chain that stretches from exotic memory stacks to glass cables, from century-old electrical manufacturing to reactor designs that don't exist yet. Every layer has winners, and most are growing at rates that rival or exceed the GPU headlines.

Broadcom's AI revenue is on track to more than triple in a year. Micron sold out its most advanced memory before it was produced. NVIDIA invested $2 billion into a laser company. Meta committed $6 billion for glass cables — then Amazon signed its own. Vertiv's backlog has more than doubled to over $15 billion. Caterpillar's generator division grew 44% without touching a chip. GE Vernova booked more data-center orders in one quarter than in all of the prior year. And the compute layer at the top of the stack is no longer a monopoly: AMD's Helios, Google's Ironwood TPU, Cerebras's wafer-scale engines, and AMD-powered neoclouds like TensorWave are all now shipping at scale.

The whole thing is being built — and occasionally slowed — by shortages of steel wound into transformers, concrete poured by specialized crews, land near substations, and gas turbines that take years to manufacture. This is what a real infrastructure build-out looks like. It's not clean, it's not orderly, the lead times are absurd, and the fortunes being made are not all in the places everyone is looking. They never are.