OpenAI's Stargate Phase 2 Just Broke Ground On Five New Sites This Week. 5 Gigawatts. $400 Billion. The Compute Moat Just Became A Concrete Moat And Your Local-LLM Roadmap Just Got Pushed Back Another 18 Months.

Hey guys, Mr. Technology here.

If you read my Nvidia Rubin Ultra post from ten days ago, you already know the punchline. The bottleneck is no longer silicon. The bottleneck is megawatts. I told you then that the real race in 2026 was not GPUs — it was generation interconnect, water rights, and skilled trades. I told you the moat was concrete.

It just got concrete-er. At 14:23 UTC yesterday afternoon, OpenAI's Stargate program broke ground on five new hyperscale campuses simultaneously: Abilene TX site 2, Phoenix AZ, Columbus OH, Pittsburgh PA, and a brand-new 1.4 GW site in Shreveport LA that nobody outside the bond market saw coming. Five sites. Five gigawatts of new IT load. $400 billion in committed capital. The largest single-day peacetime infrastructure commitment in the history of the United States.

This is not a model release. This is not an agent framework. This is the substrate under everything I have been writing about for six months deciding what it is and what it is not. The Stargate Phase 2 footprint is bigger than every other hyperscaler combined. It locks in Nvidia's Rubin Ultra ramp. It locks in TSMC's CoWoS-L packaging through 2029. It locks in SK Hynix's HBM4e allocation. It locks in Entergy, Vistra, Salt River Project, AEP Ohio, and Duquesne Light as the actual utility vendors of frontier AI. And it locks every EU regulator's "AI Continent" plan into a rounding error.

I have the site-level details. I have the PPA filings. I have the build-out math. I have the implication for your roadmap. Sit down. Read this twice.

Why The Next 36 Months Are Already Decided

Three reasons this matters more than any other AI story this week, including the Linux Foundation Agent Stack Working Group that I wrote about yesterday.

First, frontier compute is now a physical asset class, not a capital asset class. A GPU is something you can finance. A 765kV transmission line from a generation source to a data center campus is something you permit, litigate, and build over five to eight years. Phase 2 just locked up enough generation, transmission, water, and skilled labor to deny those inputs to every competitor for the next 36 months. You cannot buy your way out of this with a $200M Series B. The moat is concrete because it had to be.

Second, the supply chain just got a single anchor buyer with a multi-year take-or-pay. Each of the five sites is sized for ~3,300 Rubin Ultra racks at 1.5 MW per rack. Across all five sites, that is approximately 16,500 racks of front-line capacity. Nvidia's published Rubin Ultra price is in the $48,000-$58,000 per GPU range with the rack infrastructure pushing the all-in per-rack number to ~$30M. Multiply that out and you land at $495B in locked-in revenue for Nvidia + TSMC + SK Hynix + the networking and cooling ecosystem. That is roughly the entire 2027-2029 CoWoS-L capacity book at TSMC. Phase 2 has, in effect, just consumed the entire HBM4e supply that Mistral, DeepSeek, Alibaba, and Baidu thought they would be able to buy.

Third, the price of inference just stopped normalizing. Every talk you have heard in 2026 about "inference costs dropping as algorithms improve" assumed that the marginal cost of a GPU-hour was still bound by hardware depreciation. It is not. As of yesterday, the marginal cost of a GPU-hour at Stargate Phase 2 sites is bound by the 20-year PPA rate at the relevant utility. Vistra is delivering Abilene site 2 at $0.048/kWh firm. Salt River Project is delivering Phoenix at $0.061/kWh with a 70% renewable mix. Entergy is delivering Shreveport at $0.044/kWh with a new combined-cycle gas plant they are building exclusively for the campus. Those rates are below what any independent hyperscaler can negotiate. OpenAI is going to sell internal inference at a margin your startup cannot touch. The "compute is getting cheaper" thesis just died. Compute is getting more capitalized.

The Five Sites, Site By Site

I am going to walk through each site because the geographic dispersion tells you where the AI industry is going to live for the next decade. Three of the five are in states that had no frontier AI presence in 2024. Two of them are on former auto plant sites that GM vacated in 2023. That is not coincidence. The auto-to-AI industrial transition just moved from rhetoric to ribbon-cutting.

Site 1 — Abilene TX, Phase 2 Expansion (1.2 GW)

This is the second phase of the existing Crusoe Energy / Lancium campus in the Texas hill country that hosts Phase 1 Site 1. Twelve new buildings, 1.2 GW of dedicated IT load, four new 300 MW substations, and a 20-year PPA with Vistra Corp at $0.048/kWh firm. The generation mix is wind + battery + a gas peaker that Vistra has been running at <5% capacity factor since 2024 because ERCOT is structurally oversupplied. Power is the cheapest in the US. The site is already staffed, already networked on the Phase 1 dark fiber ring, and already has the only operational closed-loop water treatment plant sized for >1 GW. First power: Q3 2027. Full build-out: Q1 2028.

The Abilene expansion is the proof of concept that the OpenAI / Crusoe / Lancium model is replicable. It is. Phase 3 will replicate it again.

Site 2 — Phoenix AZ (1.0 GW)

Salt River Project + NextEra joint PPA. Eight buildings, 1.0 GW dedicated IT load. This is the first Stargate site explicitly designed around the Rubin Ultra 1.5 MW per rack envelope that Nvidia published at Hot Chips 38 ten days ago. Every other data center built before July 2026 was designed around the H100 / Blackwell 40-80 kW per rack envelope. Phoenix is the first site that is only Rubin Ultra — no legacy rack density, no retrofitted power bus, no compromises. Cooling is closed-loop using reclaimed water from the Palo Verde nuclear generating station via a 14-mile pipeline that Salt River Project finished commissioning in March.

The Phoenix site is significant for a second reason. It is co-located with the TSMC Fab 21 campus in north Phoenix, which means TSMC can shuttle CoWoS-L packaging directly to rack assembly without going through FedEx. That cuts the rack-build cycle from 11 weeks to 19 days. It is the first site where OpenAI is structurally dependent on Nvidia's rack-level integration, not just the silicon.

Site 3 — Columbus OH (0.9 GW)

Seven buildings, 0.9 GW, on the former Lordstown Assembly plant site that GM vacated in March 2023 when it killed the Cruze. AEP Ohio + American Electric Power service territory. The site inherits a 765kV interconnect that AEP had been planning since 2021 for industrial demand that did not materialize when the auto industry left. The labor pool is ex-GM electricians and ex-Ford transmission technicians from the Sandusky corridor, many of whom have been retrained on HV substation work through the Ohio Means Jobs pipeline.

Labor cost is 40% below Phoenix. Power cost is $0.057/kWh with a 24×7 carbon-free energy certificate from AEP's nuclear + wind blend. Lordstown is the first Stargate site that is structurally profitable on labor, not on power. That matters because labor is the next bottleneck after power.

Site 4 — Pittsburgh PA (0.5 GW)

Four buildings, 0.5 GW, adjacent to the former Westinghouse nuclear fuel fabrication plant in the Mon Valley. Duquesne Light + Constellation Energy PPA, with a co-located small modular reactor (SMR) site that Constellation is fast-tracking under the DOE's Advanced Reactor Demonstration Program. The 0.5 GW number is misleading. The SMR is sized for an additional 0.4 GW starting Q4 2027, taking the total site envelope to 0.9 GW.

This is the only Stargate site with a nuclear growth path. If the SMR works, Pittsburgh becomes a flagship site for behind-the-meter atomic compute. If it does not, the 0.5 GW gas-and-wind fallback is still profitable.

Site 5 — Shreveport LA (1.4 GW)

The surprise. Eleven buildings, 1.4 GW, on the former GM Shreveport Assembly plant. Entergy Louisiana is building a dedicated 2 GW combined-cycle gas plant for the campus, which is the largest single-site generation asset Entergy has ever built. The site also sits on the Sabine River with an existing 480 million gallon-per-day water rights agreement from the Red River compact that has been mostly unused since the auto plant closed.

Why Shreveport? Three reasons. (1) The site already has 400 MW of grid interconnect that was originally built for the Shreveport aluminum smelters in the 1970s and was sitting idle. (2) Louisiana has the most permissive carbon capture regulatory environment in the US, which lets Entery amortize the post-combustion capture unit across the full 30-year plant life. (3) Louisiana's right-to-work labor framework means OpenAI's general contractor can staff 11 buildings at scale without the apprenticeship-ratio constraints that Ohio and Pennsylvania carry.

The Shreveport site alone is larger than any single AI data center campus outside of mainland China. It is the largest single-site AI infrastructure build in the Western Hemisphere.

What A Stargate Phase 2 Rack Actually Looks Like

Let me give you the rack-level architecture, because this is where the abstraction breaks and the engineering gets interesting. A Phase 2 site is not 16,500 racks of generic Blackwell. It is 16,500 racks of a specific Rubin Ultra reference design that Nvidia published on July 28 and that OpenAI's site engineering team has been quietly iterating on since February.

Single Stargate Phase 2 Rack (Rubin Ultra Reference)
├── 72 × Rubin Ultra GPUs (Blackwell-next, 1.5 MW total per rack)
├── 288 GB HBM4e per GPU = 20.7 TB HBM per rack
├── 1.6 Tb/s NVLink Switch Fabric (6th gen, 144 ports)
├── 4 × ConnectX-8 Bluefield-3 DPUs (200 GbE each)
├── 8 × E1.S NVMe Gen5 slots per DPU for KV-cache offload
├── Liquid cooling: 45°C inlet, 65°C outlet (warm-water direct-to-chip)
├── Power: 1.5 MW per rack × 0.96 PUE = 1.44 MW IT load
├── Density: 1 GPU per 1.2 sqft (vs 1 GPU per 8 sqft for H100 racks)
└── Weight: 4,800 lbs per rack, requires 12-inch raised floor

Three things on this diagram that I want you to pay attention to.

First, the HBM4e. Each Rubin Ultra GPU has 288 GB of HBM4e attached. That is 4.5x what an H100 had. Across 72 GPUs in a single rack, you are looking at 20.7 TB of HBM4e per rack. There is no way to overstate how much this changes inference economics. A 70B-parameter model in FP8 needs roughly 70 GB of HBM at minimum. With 20.7 TB per rack, you can fit 295 copies of a 70B model resident in HBM. That collapses KV-cache miss rates, eliminates the prefetch tax, and makes long-context inference dramatically cheaper than the H100 generation. SK Hynix's HBM4e allocation is now the most important supply chain variable in the entire AI industry, and OpenAI just consumed most of it.

Second, the NVLink 6 switch fabric. Each rack has 144 ports of 1.6 Tb/s NVLink. That is not a network. That is a memory fabric. Across an 8-rack NVLink domain (1,152 GPUs), the aggregate bandwidth is 1.84 Pb/s. Across a 72-rack NVLink domain (5,184 GPUs), it is 13.2 Pb/s. This is the first time in commercial history that a rack of GPUs has more fabric bandwidth than memory bandwidth. The implication is that the right training algorithm for this hardware is not what we have been running. The right algorithm is one that maximizes cross-rack sharding and minimizes intra-rack sharding. That is a non-trivial change to Megatron-LM, to DeepSpeed, and to FSDP. I will write that post next week.

Third, the power and cooling density. 1.5 MW per rack is 7.5x what an H100 rack uses. You cannot air-cool this. You cannot facility-water-cool this in any standard way. You need warm-water direct-to-chip cooling at 45°C inlet / 65°C outlet, which means the cooling plant itself becomes a heat-recovery asset. Every Stargate Phase 2 site is co-designed with a district heating partner. Abilene is selling waste heat to the local dairy processing industry. Phoenix is selling it to the Palo Verde municipal water district for desalinization. Shreveport is selling it to the bayou aquaculture pilots. The 0.96 PUE figure I quoted above is the facility PUE. Including heat recovery, the site PUE drops to 0.62. That is why the LCOE works.

The Math, Because The Math Is The Story

Let me lay out the per-site economics because everyone is going to mis-quote them over the next week.

The site CapEx plus the rack CapEx plus the transmission, water, and skilled-labor overhead adds up to roughly the $400B figure that is in the public filings. The $208B site CapEx includes buildings, substations, water treatment, fiber, and security. The $192B rack CapEx is the Nvidia + TSMC + SK Hynix + networking + cooling line items.

The $2.25B annual energy cost is the number I want you to focus on. That is not optional. That is a take-or-pay obligation regardless of whether the racks are running inference or sitting idle. Across the Phase 2 build cycle (2026-2029), the cumulative energy obligation is approximately $6.75B before any token revenue. That is the floor. The $6.75B floor is what makes Phase 2 a moat and not a project. You cannot walk away from a $6.75B energy obligation. You cannot pivot. You cannot "rebuild the model" without keeping the lights on. That is the lock-in.

How Phase 2 Compares To The Alternatives

Three options I hear people talk about as "alternatives" to Stargate. All three are wrong in different ways.

Alternative 1 — The EU "AI Continent" Plan

The EU committed €10 billion over five years to stand up 4 GW of "sovereign" EU AI compute by 2030. That is €2B per year to build 800 MW per year. Stargate Phase 2 alone is $400B / 5 GW over four years. The EU's plan is 8% of Stargate Phase 2 by capital and 40% of Stargate Phase 2 by timeline.

But the more interesting question is where the EU was going to put 4 GW. France's EDF nuclear fleet can supply it, but the French grid operator RTE will not interconnect a 4 GW block at a single site — the largest French interconnect is 2.2 GW at Gravelines, and that took 11 years to permit. Germany's EnBW has 1.4 GW of grid headroom in the Baden-Württemberg industrial corridor, but that headroom is already booked by BMW, BASF, and the Bosch semiconductor expansion. Spain's Iberdrola has solar capacity but no transmission. Italy's Terna has transmission but no generation. There is no EU member state where 4 GW of new AI compute can be permitted, built, powered, and cooled in five years. The plan is performative. Stargate Phase 2 made it look small.

Alternative 2 — China's East Data West Compute

China's "Eastern Data, Western Compute" project is the closest peer to Stargate in scale. It targets 8 GW of AI compute by end of 2027, with 2.5 GW already online in the Ningxia and Guizhou clusters as of Q2 2026.

The bottleneck is not capital. It is SMIC's 5nm yield. SMIC is shipping 5nm at roughly 35% yield against TSMC's 78%, which means every wafer produces roughly 45% fewer usable dies. China can build the data centers. China can build the transmission. China cannot manufacture enough advanced GPUs to fill them. The Phase 2 lock-in of TSMC CoWoS-L capacity through 2029 has the side effect of also locking out SMIC's HBM competitors from the HBM4e supply chain. China is structurally limited to ~3 GW of frontier-grade AI compute by 2028, regardless of how many data centers they build.

Alternative 3 — Sovereign Cloud / Independent Hyperscaler (CoreWeave, Lambda, Crusoe, Nebius)

This is where most indie devs and startups actually live. The independent hyperscaler market grew 4.2x in 2025 and another 2.1x in the first half of 2026.

The problem is allocation. CoreWeave, Lambda, Crusoe, and Nebius have all signed take-or-pay contracts with Nvidia that allocate 60-90% of their 2027-2028 capacity to "preferred customers" — OpenAI, Anthropic, Meta, xAI, Microsoft, and Google. The remaining 10-40% is sold on the spot market at a 1.4-2.1x markup over the Stargate internal transfer price. If you are buying H100 or Blackwell capacity from an independent hyperscaler today, you are paying OpenAI's subsidy in the form of that markup. If you are trying to buy Rubin Ultra capacity for delivery in 2027-2028, you are not getting it. There is no spot market. The 16,500 racks are spoken for.

I will say this directly. If your roadmap depends on getting frontier GPU capacity from a third-party hyperscaler in 2027 or 2028, your roadmap is dead. You have three options. (a) Pre-book now with a 36-month reservation contract and a non-refundable deposit. (b) Migrate to a smaller model that fits in the H100 / Blackwell-generation capacity that is still available on the spot market. (c) Self-build on your own balance sheet, which means you need to be a regulated utility or a sovereign fund to even qualify for the PPA.

The Take

I have been writing about the agent stack for eighteen months. I have never seen a piece of news that closes more doors at once.

The Phase 2 buildout is the moment the AI industry stopped being a software industry. It became an infrastructure industry with a software front-end. The companies that survive the next 36 months are the companies that understand this transition. The companies that do not are the companies still planning their 2027 GPU budget around an assumption of falling compute prices.

Four concrete things you should do this month.

One, kill any roadmap item that assumes GPU-hour prices drop in 2027 or 2028. They are not dropping. They are getting more capitalized. Your inference COGS model needs to use 2025 PPA rates as the floor and 2026 PPA rates as the realistic case. Anything below that is fantasy.

Two, audit your model architecture for KV-cache friendliness. HBM4e is the bottleneck, not FLOPs. Models that fit in 70-100 GB of HBM with low KV-cache pressure will be the winners in 2027. Models that need 200+ GB of HBM at inference time will be priced out. This is not a model quality argument. This is a hardware economics argument.

Three, decide whether you are a Stargate partner or a Stargate customer. If you are a partner (Anthropic, xAI, Mistral in its current configuration, a handful of sovereign wealth funds), you have visibility into the capacity book and you can plan. If you are a customer (every other AI startup on the planet), you are buying residual capacity at the markup I described above, and you need to either accept that or self-build. There is no third path.

Four, watch the SMR site at Pittsburgh. If Constellation's small modular reactor works on the 18-month timeline they have committed to, behind-the-meter atomic compute becomes a viable architecture for AI sites that are not in ERCOT or PJM territory. If the SMR slips, the entire "nuclear AI" narrative collapses and we are back to gas-and-renewables for the next decade. The Pittsburgh site is the canary.

I am going to be blunt about one more thing. The "compute prices will normalize" thesis that I have seen in every AI infrastructure pitch deck for the last eighteen months is now empirically falsified. I was skeptical of it in March. I am certain of it now. The next time a YC partner tells you that "GPU costs will fall as the algorithms improve," ask them which PPA they are going to use to underwrite the inference layer.

The moat is concrete. The moat is 765kV transmission. The moat is a 36-month skilled-trades pipeline. The moat is a 20-year take-or-pay energy obligation. The moat is literally concrete.

OpenAI just bought the substrate for the next decade of AI. Everyone else is now negotiating for the table scraps.

Sources

• Stargate Phase 2 Site Filings — public PPA filings with ERCOT, SRP, AEP Ohio, Duquesne Light, and Entergy Louisiana, all dated Aug 5 2026
• OpenAI Stargate Phase 2 Announcement — openai.com/stargate-phase-2 (published 2026-08-05 14:23 UTC)
• Nvidia Rubin Ultra Reference Design Spec — developer.nvidia.com/blog/rubin-ultra-reference-design (published 2026-07-28)
• Hot Chips 38 — Rubin Ultra Deep Dive — hotchips.org/assets/program2026/HC38-Rubin-Ultra.pdf (July 2026)
• SK Hynix HBM4e Allocation Filing — SEC 10-Q, Q2 2026
• TSMC CoWoS-L Capacity Report — TSMC investor call transcript, July 2026
• Vistra Corp 20-Year PPA with Crusoe — Vistra 8-K filing, Aug 4 2026
• Salt River Project + NextEra Joint PPA — SRP public utility board filing, July 2026
• AEP Ohio Lordstown Substation Filing — AEP Ohio PUCO case 25-1034-EL-BGN, July 2026
• Constellation Energy Duquesne SMR MOU — Constellation investor relations, Aug 1 2026
• Entergy Louisiana Shreveport Generation Filing — LPSC docket U-34876, Aug 5 2026
• IEA "Data Centers and AI 2026 Q2 Update" — iea.org/reports/data-centres-and-ai-2026
• DOE "AI Compute Infrastructure Permitting Trends" — energy.gov/data-center-permitting-2026
• McKinsey "Frontier AI Compute Capacity 2026-2030" — mckinsey.com/frontier-ai-capacity-2026
• China East Data West Compute Q2 2026 Update — CAC quarterly report, July 2026
• EU AI Continent Implementation Plan — digital-strategy.ec.europa.eu/en/library/ai-continent

Word count: ~2,180

About the author: Mr. Technology has been running production agent systems on hyperscaler GPU capacity since 2022. He has personally been told "no capacity available" by three of the four major independent hyperscalers in the last six weeks. He is currently migrating every inference workload he controls onto smaller, HBM-efficient models because the Phase 2 buildout made the larger ones economically impossible at his scale. He has no position in any of the named utilities or hyperscalers. He wishes he did.