Sunday Essay — The Substrate Bill Comes Due

For a decade the conversation about artificial intelligence has been almost entirely about the model. Which architecture, which dataset, which loss, which alignment trick. The thing the model runs on was treated as a logistics problem you handed to the cloud team. Look at the announcements of the last six weeks and that conversation has changed shape, even if the headlines have not caught up. The substrate is the story now, and the bill it has been quietly accumulating has come due.

What changed is not a single announcement but a stack of them, each modest on its own, devastating together. The rack-level power envelope crossed a threshold above which the building it sits in has to be rebuilt. The memory under the GPU sold out for the year before the GPU itself shipped. The grid connection to the building cannot be procured inside an investment horizon. The cooling system stopped being optional. And the dominant interconnect quietly stopped being a standard and became, again, a vendor product.

I want to walk through what those five facts mean for anyone making a decision about AI infrastructure this quarter, because the press releases are still graded as wins. They are wins. They are also, taken together, a description of a system that has lost most of its slack.

the rack is now a power plant

On 1 June 2026, CoreWeave brought up the first NVIDIA Vera Rubin NVL72 at its Livingston, New Jersey site, delivered by Dell on a PowerEdge XE9812. Bring-up to production took under six and a half hours from delivery, a logistics milestone CoreWeave was right to claim (CoreWeave). Seventy-two Rubin GPUs, thirty-six Vera CPUs, 3.6 exaFLOPS at FP4, 260 TB/s of NVLink 6 fabric bandwidth.

The number the press release buries, and the one a CFO needs to read first, is the rack's power envelope: 190 to 230 kilowatts (TechTimes). A single NVL72 draws what roughly 150 American homes do in mid-summer. There is no air-cooled version. NVIDIA shipped this platform with 100 percent direct-to-chip liquid cooling as a hard requirement, not a deployment option, and the cooling retrofit alone runs $500 to $1,500 per kilowatt of capacity — call it $60,000 to $195,000 per rack before anyone touches the electrical plant.

I am not flagging this as a complaint about NVIDIA. The engineering is doing what the physics demands. I am flagging that the unit being deployed is not what most enterprise infrastructure teams still draw on their slides. The slide shows a server. The reality is a rack that draws a quarter of a megawatt and refuses to run unless the building's chilled-water loop is the right size and the regional substation can wheel it the power.

That is the easy part. The hard part is the power itself.

what 800 volts is actually saying

At GTC 2026 in March, NVIDIA and Texas Instruments unveiled a complete 800-volt direct-current power architecture for AI data centers (Texas Instruments, 16 March 2026; NVIDIA developer blog). Vertiv's 800VDC ecosystem, built against the Vera Rubin Ultra Kyber platform, ships commercially in the second half of this year. CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, and Together AI are designing new facilities directly to 800VDC distribution. SemiAnalysis has been tracking the whole transition under its "800VDC" frame (SemiAnalysis).

A few percent efficiency gain over the 54-volt status quo is the cover story. The actual story is that you cannot deliver several megawatts of power at the rack level using existing low-voltage bus architectures without unreasonable amounts of copper. The industry has reached for the only solution the electric-vehicle world has already debugged: high-voltage DC. Total incremental capacity that will sit on 800VDC by 2030 is now forecast at roughly 39 GW. That is a continent-scale build, and it requires a power-electronics supply chain that does not yet exist at the volumes implied.

Two things follow. First, anybody planning a facility for a 2027 ready date and not specifying 800VDC at the rack is shopping for a building that will need to be retrofitted before it depreciates. Second, the buyers of solid-state transformers, sidecar racks, and 12-volt-output power-shelf modules are about to discover that their procurement cycle is the new bottleneck. The chip is no longer the long-lead item. The conversion stage is.

the memory tax

Here is the number that has been quietly rewiring enterprise IT capex without making the cover. DRAM prices surged ninety percent in the first quarter of 2026 alone, on top of price hikes already booked through 2025, and S&P Global, IEEE Spectrum and others have all walked through the mechanics (IEEE Spectrum; S&P Global; The Register, 2 June 2026). Gartner's figure, picked up across enterprise buyers, is that memory prices rose between 50 and 200 percent in the first half of 2026, pushing server costs up by more than 125 percent. OEM quotes now expire in a week. The contract you signed in March is not the contract you executed in May.

The mechanism is not mysterious. The three memory manufacturers — SK Hynix, Samsung, Micron — have reallocated cleanroom capacity from commodity DRAM to high-bandwidth memory destined for the Rubin and MI400 lines. SK Hynix's CFO told the market last quarter that the company's entire 2026 HBM supply was already sold (TrendForce, March 2026). Samsung began HBM4 shipments in February 2026. Micron's slice is reportedly being directed toward Rubin CPX, the inference-tier variant, rather than the flagship. Those vendor allocation figures deserve a pinch of salt — they are reported by parties with an interest in the headline — but the direction is consistent across independent trackers.

The wafers go somewhere. They are not going into the LPDDR module in your finance team's laptops or the SODIMM in your call-center fleet. Consumer and enterprise client tiers are paying the substrate's bill on behalf of the AI tier, and the math is going to show up in capex plans for the next two budget cycles. Most forecasts compiled in early June see prices climbing through the rest of 2026; the earliest plausible normalisation window stretches into 2027 or 2028. If a CFO is working off a TCO model signed off in October 2025, that model is wrong, and not by a little.

The interpretation that this means a generalised AI-spend bubble is a leap I would not yet make — much of the demand is contracted to genuine deployments, not speculative buys. But the consequence is unambiguous: any model TCO calculation that does not separately track memory at current strip pricing is producing a number too low by tens of percent.

the four-year wait

The grid is older than the chips. In the United States, transmission interconnection queues for new data-center loads are now multi-year; large studies routinely cite four years from request to energisation, and the recent industry literature is consistent on this even where the headlines differ on cause (Brookings). Gartner's projection that power shortages will restrict forty percent of AI data centers by 2027 is the figure the industry has finally accepted, even where it does not like the implication (Data Center Knowledge).

Two responses are emerging, and they are uneven in quality. The first is bring-your-own-power: behind-the-meter generation, gas turbines on site, direct deals with merchant generators. This works in a narrow sense, until it is asked to scale or to pass an environmental review. The second is sovereign and quasi-sovereign siting — moving the gigafactory to where the megawatts are, or where the regulator has decided to clear them. That is the path the EU is now testing.

the european substrate bet

The EU AI Gigafactories initiative, anchored on a €20 billion InvestAI facility, intends to seed up to five sovereign AI factories, each meant to host more than 100,000 advanced AI processors (European Commission; Polytechnique Insights). On 5 June 2026 the AION consortium — Ardian, Artefact, Bull, Capgemini, the EDF group, the iliad Group, Orange and Scaleway — filed a French candidate bid (StorageNewsletter). On 17 June 2026, Bull and Foxconn announced they will manufacture and assemble the Vera Rubin NVL72 platform in Europe, the first time Rubin-class hardware will be put together on European soil (TechTimes, VivaTech coverage).

I have been more skeptical of the European sovereignty pitch than is fashionable. The candid summary is that the program has slipped, the bidding remains in a preparatory phase that was originally meant to close in late 2025, and several analysts now openly describe the €20 billion as a sovereignty mirage if the dependency on NVIDIA silicon and HBM4 supply is not addressed (CTOL Digital). Five gigafactories at 100,000 GPUs each is half a million GPUs. Individual U.S. hyperscalers have publicly committed to more than that in single-facility builds. Numerical sovereignty, this is not.

What I will defend is the manufacturing-on-soil decision. Bull and Foxconn assembling NVL72 in Europe is the kind of mid-stack control that survives a political tariff turn. It is the substrate bet that does not need EU silicon to be useful — only EU integration. Geopatriation, the term I have been using to describe the localisation of AI workloads under sovereign pressure, is best understood not as silicon repatriation but integration repatriation: the rack is assembled here, the cooling loop runs on water that did not cross a border, the power is contracted under EU law, the inference is served from local jurisdiction. That can be done with foreign silicon. The U.S. did the same with TSMC for two decades.

the interconnect quietly stopped being open

The compute conversation has been drifting for a year toward the interconnect — the bus that lets a hundred-GPU rack act like one machine. The official industry position is that UALink, the Ultra Accelerator Link consortium founded by AMD, Broadcom, Cisco, Google, HPE, Intel, Meta and Microsoft, would offer an open alternative to NVLink. The consortium published its first specification in May 2025, but commercial silicon is not expected until late 2026, with meaningful deployments stretching into 2027 (ABI Research).

While the consortium was writing its standard, NVIDIA built a moat with chrome on it. NVLink Fusion is the strategic answer: instead of opening NVLink as a standard, NVIDIA opened it as licensable IP and signed up custom-silicon partners. SiFive committed to NVLink Fusion for its RISC-V server line in January 2026 (The Register). Marvell took a $2 billion investment to bring XPU customers into the Fusion ecosystem (SemiWiki). Fujitsu's MONAKA CPU and Qualcomm's custom server chips are following.

The honest read: by the time UALink ships silicon at scale in 2027, NVLink will be on its seventh generation, with custom-silicon partners building chips around it. The open alternative arrives late, against an incumbent that has out-shipped it by an order of magnitude. I would bet against UALink v2 commercial volume reaching meaningful share before 2028, and I would bet against the public framing — "open" — surviving contact with how this market actually procures. Open standards survive on a healthy commodity assumption underneath. The commodity is gone.

what the model layer learned

There is an underappreciated symmetry on the software side. The inference stack — the engines that take a trained model and serve it — has been consolidating in a way that mirrors the substrate. Benchmarking from March 2026 on Llama-3.3-70B-Instruct in FP8 puts TensorRT-LLM in front on raw throughput, SGLang sitting between TensorRT-LLM and vLLM, and vLLM still close enough that operational simplicity wins for most teams (Spheron, March 2026). On smaller models, the SGLang advantage on shared-prefix workloads can hit 29 percent over vLLM. None of these gaps is dramatic enough to change the team you hire. All of them are dramatic enough to change per-token cost on an 18-month contract.

The pattern is the same as on the substrate. The lead vendor option is faster and harder to replace. The open option is good enough for the median case and getting better. The buyer's discipline is to know which of those two descriptions applies to their workload mix and budget posture, and to refuse to be flattered into the wrong choice by a vendor benchmark.

AMD's quieter bet

AMD's MI400 lineup, unveiled at CES 2026, lands a competitive spec sheet: 432 GB of HBM4, 19.6 TB/s of memory bandwidth, 40 PFLOPS of FP4 (Data Center Dynamics; TechPowerUp). The MI455X targets the flagship tier; the MI450 is the volume part; the MI430X is explicitly pitched at sovereign AI buyers. AMD has moved to an annual refresh cadence and confirmed the MI500 for 2027.

I would not yet write AMD out. The spec sheet is competitive. What AMD does not yet have is the rack-level systems story — Helios, the answer to NVL72, is still earlier in customer deployment. The interconnect ecosystem around AMD is anchored on UALink, which means AMD's rack-scale story is gated by the same delay UALink is. The question for any large enterprise buyer this quarter is not whether to wait for AMD parity — by the time UALink ships at volume, Rubin Ultra is in the field — but whether having an AMD second-source line at moderate scale is worth the operational complexity, given the concentration risk a single-supplier roadmap implies.

That is a judgement call. If I were on the board of a Tier-1 model trainer right now, I would push to keep a 10–20 percent AMD line in the procurement plan, knowing the performance per dollar will not match NVL72, on the explicit theory that single-supplier concentration risk is cheaper to hedge against than it is to mitigate later. If I were on the board of an enterprise that buys inference capacity rather than building it, I would not pay the complexity cost.

the takeaway, ungarnished

The substrate has caught up with the model, and it changed colour while doing so. It is now a power-engineering problem with a memory-supply problem on top, an interconnect lock-in problem underneath, and an integration problem the EU is testing whether it can solve in its own jurisdiction. None of this is exciting in the way a new model release is exciting. All of it determines whether a new model release reaches customers under economics that work.

A CTO at a large bank, manufacturer, or pharmaceutical company in Europe this month has four things to do. Assume the 2026 server refresh has been re-priced by memory, and require the OEM to write the strip-price exposure into the contract. Refuse to sign any new facility deal where the rack design assumes less than 200 kW or omits 800VDC. Ask the cloud provider, in writing, how much of the inference capacity being paid for is on a contracted power line versus a queued one. Model what an AMD second-source position would cost, even if the procurement team's answer ends up being no.

The fashionable conversation in 2026 is still about which model. The honest conversation is about which substrate. The two are no longer separable, and pretending otherwise is the most expensive mistake on the table.

---

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.