The Gas-to-AI Tracker separates press-release megawatts from bankable molecules — and shows who controls the bottlenecks between gas supply and AI compute.
The market is counting megawatts. The contracts count molecules. That unit error is changing who is winning the AI gas boom.
In the press cycle, the Permian won April: Chevron–Microsoft, VoltaGrid–Oracle, Project Jupiter. Gigawatt-scale. Attention-getting. In the contract file, Appalachia is still ahead. The difference is not ideology. It is accounting. Gigawatts are not gas contracts. Once the April MW announcements convert into modeled gas burn at the canonical 75% capacity factor, Chevron + VoltaGrid together account for ~586 MMcf/d combined — not the 4.7 Bcf/d a casual read of the press releases would suggest. Net result: Appalachia still leads Permian by 2.6× on producer-attributed exposure. EQT alone is 38% of Appalachia + Permian. Utilities sign the contracts. Hyperscalers pay the bills. The companies that control more than one link in the chain — molecule → pipe → turbine → utility wrapper → hyperscaler credit — are not just supplying AI. They are becoming the infrastructure layer beneath it.
21.11 Bcf/d is the size of the chase. 4.89 Bcf/d is the part already underwritten by signed gas supply. Read the gross as the disclosed gas-to-AI contracting universe — not as gas that flows tomorrow. The investable number depends on how much legal certainty you require; every column to the right is a tighter filter on the same evidence.
The ranking flips when you count contracts.
Since December 2024, roughly twenty-one billion cubic feet a day of gas has been publicly committed to AI buildout. Sixty-one rows, fifty-six counted in the gross, five on the watchlist. The April 2026 press releases mostly say "Texas." The contracts still mostly say "Pennsylvania."
When EQT signs Duke Energy to 800 MMcf/d for ten years, EQT is putting its molecules on the line. When Kinder Morgan builds Trident, KMI is putting steel in the ground to move someone else’s molecules. Both press releases look like "Texas gas for AI." Only one of them commits Texas gas.
On the legacy non-midstream-exclusion classifier, Appalachia producers committed ~5.07 Bcf/d across fifteen rows; the Permian was ~1.94 Bcf/d. Under the stricter v3 role-aware molecule classifier (operator_role ∈ {producer-supply, integrated-supply}), Appalachia is ~3.41 Bcf/d, Permian ~2.02 Bcf/d. The Permian still leads on gross basin exposure (8.0 Bcf/d) because of pipe across thirteen. Measured EQT remains the single biggest producer at 2.65 Bcf/d, about thirty-eight percent of Appalachia-plus-Permian. Measured The story of 2025 was "one company." The story of 2026 is "the second front got loud, but the gas is still in Pennsylvania."
What changed in April 2026? Three announcements made the Permian look like it had caught up. Chevron-Microsoft signed exclusivity on a 2.5 GW West Texas gas hub. VoltaGrid disclosed a 2.3 GW Oracle master contract spanning multiple Stargate sites. Energy Transfer filed Green Chile with FERC for Project Jupiter — 388 MMcf/d for what Oracle and OpenAI describe as a $165 billion Stargate campus in New Mexico. Loud, gigawatt-scale, attention-getting. But strip out the Chevron and VoltaGrid power-capacity-as-gas-volume errors, and Permian producer commitments add up to 1.94 Bcf/d. The signatures are on power deals, not gas. Diamondback is still "in talks." Pioneer hasn’t announced anything. ConocoPhillips, Devon, EOG, Permian Resources — silent. The supply-side advantage hasn’t been matched with producer signatures yet.
The Permian pipelines are still being built the direction they were always going to be built — Trident, Desert Southwest, Copper State, now Green Chile. All announced in the last eighteen months, all designed to move Permian gas somewhere with load. The Stargate program is starting to change the "somewhere" — Chevron-Microsoft and VoltaGrid-Oracle propose to burn Permian gas in West Texas at gigawatt scale. The supply-side commitments to back those plants haven’t arrived yet, but the announcements suggest the geography of demand is shifting.
Why did Appalachia move first? Basis. Tetco M-2 traded under two dollars in 2024 Reported while Transco Zone 5 in the Carolinas was close to three. Build the pipe south, and every Bcf/d you ship is worth roughly a hundred million dollars a year in higher realizations Modeled — a number large enough to focus any CEO. The Permian doesn’t have that problem. The Permian has the opposite problem: too much associated gas relative to oil, which is how you end up flaring or paying people to take it. Until 2026, that meant the Permian exported its gas. The Chevron, VoltaGrid and Jupiter announcements suggest some of it might stay in-basin going forward — but the producer-side gas commitments to back those plants haven’t been signed yet.
Then there’s the consolidation. Williams now owns the gathering, the transport (Transco, now with SESE and NESE finally approved), the generation (Socrates, Apollo, Aquila, Socrates the Younger, plus two undisclosed $3.1B projects announced in October), and the offtake contracts to Meta. EQT — after buying Equitrans in 2024 — owns production, Mountain Valley transport, and four hyperscaler-adjacent supply deals. Chevron owns Permian gas, a 2.5 GW power hub, and an exclusivity lock with Microsoft. The companies that used to sit at separate tables are starting to look like one table. Call it the integrated gas-to-electron business. So far, it works mostly on one coast — Pennsylvania — with a Permian counterpart announced but not yet locked down on the supply side.
Watch the ranking flip.
Three different ways to ask "which basin is winning the AI gas wave." The first two — deal count and MW announced — put the Permian on top. Apply the only filter that matters (an actual signed gas contract), and the order reverses. This is the entire thesis in one chart.
If you have ninety seconds.
The core claims in order of importance. Every finding links to the evidence.
Appalachia committed; the Permian announced.
Each dot is a disclosed deal. X-axis is announcement date. Size is MMcf/d. Color is basin. Watch June–July 2025: EQT drops four deals totaling 2,645 MMcf/d in six weeks. Then March–April 2026: Chevron-Microsoft and VoltaGrid-Oracle announce gigawatt-scale Permian deals — but the gas math behind those deals shrank dramatically once unit errors were corrected.
What the shape tells you
Dec 2024 – May 2025: early, opportunistic. ExxonMobil quietly announces a CCS-enabled data-center power plant. Kinder Morgan greenlights Trident. Diamondback floats a JV. Nobody’s committed yet.
June – July 2025: the cluster. EQT signs Duke, Southern, Homer City, and Shippingport in six weeks. Williams takes Transco SESE to FID. Crusoe and Tallgrass announce Project Jade in Wyoming on July 24 — 2.7 GW of behind-the-meter gas, eventually scaling to 10 GW. EQT was the loudest voice but it wasn’t a solo act.
August 2025 – present: the Permian’s midstream response — and the first sign of Permian producers going vertical into power. Energy Transfer Desert Southwest, KMI Copper State, ET’s Cloudburst, and three more Crusoe gigawatt campuses (Abilene 2 for Microsoft, Goodnight for Google, plus Monarch in West Virginia for Microsoft via Nscale). In March 2026, Continental Resources — the largest private O&G producer in the world — announced its first-ever power plant, a 452 MW merchant generator in Pecos, Texas, co-developed with Mercuria. But the upstream supply contracts keep landing in Appalachia (PPL, Antero, Range, Williams Socrates). The pipe keeps going west to east. The signatures keep happening in Pittsburgh.
Appalachia committed. The Permian announced.
Left dot: gas the producers have actually put under contract. Right dot: pipe capacity being poured into the ground. Lines connect basins across both sides. Appalachia anchors at 5.1 Bcf/d producer supply with 400 MMcf/d in new pipe. Permian sits at 1.94 Bcf/d producer supply (down from 6.1 in v2.3.x; see methodology) with substantial new pipe capacity. The pipe story is more bidirectional than the supply story — Permian midstream is real, but the gas backing it is mostly modeled, not contracted.
What could break this
Producer volumes mix firm contracted MMcf/d (EQT–Duke, EQT–Southern, Homer City) with numbers backed out from plant size (Chevron–Microsoft, Williams Socrates). Run those plants in simple cycle and the gas burn is higher. A handful of small-cap Permian positioning deals — Coterra–CPV, Diamondback JV, Oxy Horizon — use plant-size estimates too. The Permian producer figure may be understated by 0.3–0.5 Bcf/d on the modeled side.
If you’ve got project-level data that contradicts this, we want it. hello@sunyascoop.com.
Ten names carry seventy-eight percent of the gas.
The top ten operators carry 16.5 Bcf/d of the 21.11 Bcf/d disclosed (about 78%). The bar color says whether a company is committing gas (olive) or building pipe (tan). EQT is all olive — molecules. Energy Transfer is mixed: 1.88 Bcf/d of signed integrated supply (Cloudburst, Oracle, Fermi, Entergy-Meta) and 2.69 Bcf/d of pure transport capacity (Desert Southwest, Jupiter Green Chile). The composition of each bar is the strategy.
The rest of the country will eventually get to where we are. It is a matter of physics, chemistry, and geology. Pennsylvania is the Saudi Arabia of natural gas.
What could flip this ranking
A few of these labels are shorthand. "Williams Companies" bundles four Power Innovation projects with the SESE pipeline — pull SESE out (it’s pure midstream) and Williams drops to ~492 MMcf/d of producer-equivalent supply. "Energy Transfer" is entirely midstream. The 4,565 figure is pipe capacity, not molecules committed (Cloudburst, Fermi, Jupiter, Oracle, Entergy-Meta, plus the 2.3 Bcf/d Desert Southwest expansion to APS/SRP).
One disclosed integrated deal from a Permian major — EOG, ConocoPhillips, Devon — reshuffles this chart inside a quarter. None of them have announced one yet. Ask yourself why.
The gas origin map is not the data-center siting map.
One bubble, one deal. Area is MMcf/d. Color is the producing basin. Hover for the summary, click for the full dossier with numbered sources. The cluster that forms in Appalachia and the cluster that forms in East Texas are not the same story.
What the geography hides
Bubbles sit where the gas comes from, not where it ends up. Most of the Appalachia dots deliver into the Carolinas, Georgia, or Alabama via Transco — the contract is an Appalachia contract, but the AI workload sits in the Southeast. Energy Transfer’s San Marcos bubble is even further off: plot is where the pipe ends, not where the molecules started (Permian).
If you’re reading this as a data-center siting map, you’re reading the wrong map. The electrons land in the Southeast and ERCOT. The gas comes from elsewhere.
Utilities are the contracting wrapper. Hyperscalers are the economic buyer.
Three basins on the left. Four types of buyer on the right. The widths show which pairings actually got signed — not the theoretical ones on industry conference slides, the ones with a press release attached.
Why the utility band looks the biggest
The utility flow dominates because it absorbs regulated CCGT offtake — EQT–Duke, EQT–Southern, PPL, Entergy–Meta. On paper these count as utility deals. Off paper, the end user is a hyperscaler sitting behind the meter. Reclassify by ultimate consumer and the hyperscaler flow roughly doubles.
The cleaner way to read it: utilities sign the contracts. Hyperscalers pay the bills. The utility and the grid sit between them as the regulated delivery and reliability wrapper — recovering the cost of service through rate base, not skimming a spread.
What the operators actually said.
Pulled straight from press releases, earnings calls, and SEC filings. No paraphrase, no sanitization. Filter by counterparty type or search by name. Click through to the deal.
From the Homer City Redevelopment release: the EQT–Homer City project is described as one of the largest single-site natural gas purchases in North American history, structured as an agreement in principle with first power originally expected 2027; Homer City spokesman said Apr 2026 the timing is now 2028. Homer City Redevelopment · BusinessWire, Jul 15 2025 · paraphrased from release
From the Project Jupiter announcement: Oracle, OpenAI, and Energy Transfer are partnering on a Stargate campus in New Mexico, with up to 2.45 GW of Bloom Energy fuel-cell capacity replacing earlier-planned gas turbines and diesel generators. Oracle Newsroom · Apr 27 2026 · paraphrased from release
From the VoltaGrid–Oracle release: VoltaGrid will deliver 1+ GW of behind-the-meter natural-gas-fueled power across multiple Oracle data-center sites, with Nathan Ough (CEO, VoltaGrid) describing it as the largest gas-to-compute project of its kind. Reuters / VoltaGrid · Oct 15 2025 · paraphrased from release
The scoreboard is not volume. It is certainty.
The standard read of an AI-gas tracker is "who has the most Bcf/d?" That's the wrong question. The right question is what to do with each row given how confident you are it lands. Five evidence classes, four decisions.
Most of the press lives in the bottom-right quadrant. Most of the bankable economics live in the top-right. The interesting work as an underwriter is the diagonal — Underwrite vs Watchlist — because that's where the unit error happens.
Utilities sign. Hyperscalers pay.
The contract on paper is not the contract that matters. Four roles in every gas-to-AI deal — and they almost never sit at the same desk. The reader who keeps these straight wins the underwriting question.
The asymmetric risk: the legal buyer (utility) gets cost-of-service recovery. The economic buyer (hyperscaler) keeps optionality on the campus. The asset owner takes balance-sheet duration. The risk owner — most often a ratepayer or pre-FID developer — has the least leverage and the highest exposure.
Not a basin race. A control-stack race.
The interesting question isn't "Permian or Appalachia?" It's: who controls the most scarce links in the gas-to-compute chain? Five links, five scarcities, five sets of names worth tracking. The companies that control more than one are the platforms.
The platforms are the companies stacking links. Williams stacks Movement + Permission + Credit (Meta offtake). EQT stacks Molecules + Movement (post-Equitrans) + Credit (Toby Rice's hyperscaler chain). Chevron stacks Molecules + Permission (its own JV) + Credit (Microsoft exclusivity). Everyone else is renting one layer of the stack.
Build the AI Power Stack.
Pick the hyperscaler. The five-link chain assembles from actual signed deals in the ledger: gas → pipe → generation → permission → hyperscaler credit. Get a deliverability score, the bankable gas total, and the primary constraint that breaks the chain. Same evidence as the methodology — different lens.
Score = bankable share × 0.4 + definitive-binding share × 0.3 + average power-to-compute maturity × 0.3, on a 10-point scale. Numbers update from the ledger; nothing is hardcoded per buyer.
What this means.
Same evidence, six different decisions. Each card is the call you should be making if you sit at that desk, written for the person who will read this on a phone between meetings.
Three ways I’d be wrong.
I’ve spent fourteen months building this. I think the thesis is right. But there are three specific futures in which it isn’t, and I’d rather name them than have a reader find them. Here they are, with the leading indicators I’m watching for each.
The Permian has 24,812 MW announced and only 0.98 Bcf/d signed. The structural reason isn’t a lack of gas — it’s a lack of takeaway capacity from Waha east. Today, a substantial volume of Permian associated gas is either stranded, takeaway-constrained, or price-discounted at Waha (publicly reported as flared, vented, or sold at negative basis prices). That gas can’t move to PJM or ERCOT-East datacenters because the pipes run the wrong direction. If a major eastbound expansion gets sanctioned (KMI’s Trident, ET’s Warrior, or a new Energy Transfer Permian-to-Houston line) — and if the line gets gas-supply contracts attached, not just transport capacity — Permian signed-gas could double inside 18 months.
Project Jupiter’s April 27 fuel-cell pivot — Bloom Energy replacing 2.0 GW of planned gas turbines with solid-oxide fuel cells — was the first major data-center deal to swap technology mid-build. The fuel cell still consumes gas, but at a different heat rate, with different ramp characteristics, and with implications for who supplies (compressed gas tanks vs pipeline molecules). If three more deals of this size pivot before year-end, the modeled-burn bucket revises down by 0.5–1.0 Bcf/d, and the "signed gas" figures hold but become less representative of the true demand shape.
GE Vernova / Siemens / Mitsubishi turbine slots are oversubscribed into 2028+. If lead times stretch to 2029 — already plausible — the seven biggest behind-the-meter projects can’t break ground when their gas contracts say they must. Hyperscalers go grid, gas operators eat MOU breakage costs, and the same molecules redirect to LNG export contracts where there is unlimited offtake. If turbine slot availability worsens by 6+ months between Q2 and Q4 2026, the "Appalachia signed" volumes look the same on paper but the underlying buyer reorganizes from datacenter to LNG facility — same gas, very different industrial story.
Where private capital can deploy.
Same dataset, ranked through a PE underwriting lens. Five scores per row — scarcity, bankability, deployability, moat, exit (1–5 each) — sum to a 25-point investability total. Top-ten ranks below; investment screen on the right; archetype taxonomy below that. Heuristic, not conviction-grade — equipment-slot diligence, all-in $/MWh modeling, and per-row PE archetype assignment require primary research.
The 9-filter investment screen
- Customer Investment-grade hyperscaler, utility, or contracted intermediary
- Contract Take-or-pay, availability payment, tolling, or long-term lease
- Site Controlled land, fiber, gas, water, political support
- Power Equipment slot secured, credible first-MW date
- Permits Air, water, zoning path with no fatal opposition
- Economics Clear spread between contracted revenue and delivered power cost
- Capital External equity actually needed; can buy control
- Exit Obvious buyer at lower cost of capital after de-risking
- Downside Asset retains value if AI load slips (industrial, capacity, REIT, ancillary)
Five PE archetypes the data points to
What would change our mind.
The thesis rests on specific evidence. Here is the evidence that would weaken it — and the evidence that would sharpen it. Skepticism as a feature, not a defensive footnote.
- A Permian major (EOG, Devon, Conoco, Oxy, Marathon) signs a named, binding producer–hyperscaler gas-supply contract.
- OEM turbine delays push first-power dates past 2028 across multiple top-10 sites.
- Fuel cells (Bloom, SOFC) capture >1 Bcf/d of what we modeled as gas-turbine burn.
- Hyperscaler capex slows materially (combined Big-5 below $300B by 2027).
- Utility procurement out-paces behind-the-meter buildouts on the marginal new project.
- EQT's Duke or Southern volumes get reassigned to non-AI industrial load.
- More named producer→hyperscaler contracts in Appalachia (Range, Antero, CNX, Coterra Marcellus).
- Project-specific turbine orders confirmed at Homer City, Shippingport, Hyperion.
- PUC filings that explicitly allocate AI-driven costs to hyperscaler tariffs (not ratepayers).
- Air permits issued and first-power dates confirmed at top-10 sites.
- Contract language naming gas explicitly (not "fuel" or "thermal").
- Tetco M-2 to Transco Zone 5 basis spread widening — pull on EQT signing.
What survives your skepticism?
Flip the toggles below. Each one is the meanest version of a critique we’ve fielded — from analysts, short-sellers, the skeptics in our DMs. The headline number recomputes from the deals.json source. Pick a preset, or compose your own worldview.
- Double counting. If EQT’s Duke/Southern volumes are transported by Williams SESE, they show up in both producer supply and midstream capacity. The net 19.66 Bcf/d figure backs out the largest identified overlap; others may exist. If the true overlap is 3+ Bcf/d, Appalachia’s producer lead narrows.
- Deal non-completion. Positioning rows and FERC filings can slip. Chevron-Microsoft is signed; Project Jupiter is filed but not in service. Volumes that never flow can’t power anything.
- Turbine delays. Gas-fired capacity needs GE Vernova / Siemens / Mitsubishi turbines, and order books are stretched into 2028+. A binding equipment constraint would slow every behind-the-meter project whether the gas is under contract or not.
- Modeled-from-MW volumes are conservative. Modeled-burn rows total ~3.58 Bcf/d at P50 (P10 ~2.64; P90 ~3.43; earlier 3.25 figure was a different P-point); all modeled-from-MW rows including signed-modeled and positioning total 5.04 Bcf/d using heat-rate × capacity-factor (Chevron-MSFT and VoltaGrid use 85% CF (the AI-baseload exception); utility/IPP rows use the canonical 75% CF default per the methodology). If AI loads actually run baseload (85%+ CF), the utility/IPP rows are undercounted by up to 50%. The retracted Chevron-MSFT and VoltaGrid envelope figures (previously 2,500 and 2,200 MMcf/d) are no longer in the totals — those rows now sit at 311 and 327 MMcf/d respectively.
- Permitting and water constraints. Air permits for new gas generation are slow in PJM and contested in ERCOT. Water for evaporative cooling is a binding constraint in the Permian. Either could push sites off gas and back onto the grid.
- Hyperscaler demand slows. The thesis assumes sustained AI capex. A material slowdown in hyperscaler build-out — whether from training-to-inference rotation, model efficiency gains, or macro tightening — would leave contracted gas without a buyer.
- Load shifts back to grid procurement. Utility PPA structures with hyperscaler cost recovery might prove more economic than behind-the-meter at scale. If Entergy-Meta is the template rather than Homer City, the molecule-owner thesis weakens.
- Named producer–hyperscaler contracts from the Permian majors. EOG, ConocoPhillips, Devon, Marathon, Oxy. None have disclosed DC-linked supply deals yet. The first one signals that Chevron-Microsoft wasn’t a one-off.
- Turbine orders tied to named projects. GE Vernova, Siemens, or Mitsubishi announcing backlog pointed at specific AI campuses.
- Air permit approvals for behind-the-meter sites. Homer City, Shippingport, Fermi HyperGrid, Monarch. Approvals move these from positioning to binding.
- Utility filings with hyperscaler cost recovery. PUC dockets where regulators explicitly allocate AI-driven capex to hyperscaler ratepayers. Validates the utility-as-contracting-wrapper read.
- Producer rate-base disclosure in integrated JVs. Chevron-Microsoft and similar structures are reported at the JV level but not segmented. Segment disclosure would show whether the producer is the operator or the financier.
- Hyperscaler PPAs with explicit gas language. Most hyperscaler PPAs are gas-agnostic. Named gas commitments in the contract body (not just the press release) confirm the molecule dependency.
- Basin differential movement. Tetco M-2 and Transco Zone 5 pricing. If Appalachia basis tightens as new takeaway arrives, it confirms that the 2025 Appalachia wave was basis-driven, not structural.
All 61 rows · 56 counted, 5 watchlist. Every source.
The whole dataset, one row per deal. Click any row to open the dossier — structure, counterparty terms, quoted language, numbered sources. Filter by counterparty type. Search any field. If you spot an error, tell us.
| Date | Operator | Counterparty | Basin | MMcf/d | Decision | Status | Evidence | Cites |
|---|---|---|---|---|---|---|---|---|
| Ledger loads from deals.json · 56 counted rows · sortable, filterable · requires JavaScript. Download raw JSON → | ||||||||
Show your work.
Sorted newest to oldest. primary is SEC filings and company IR. trade is NGI, S&P Global, Hart Energy, DCD. analysis is East Daley, RBN, Zacks. corporate is company-adjacent content — press releases, investor decks, IR pages. Click anything; it opens in a new tab.
What 21.11 Bcf/d is — and is not.
What’s in scope
Any disclosed project that links a US oil & gas operator to data-center gas supply. Supply agreements. JV or equity stakes in power projects serving DCs. Midstream pipelines with a named DC endpoint. Co-located gas plants where the operator supplies molecules. Positioning announcements count, but they’re flagged as positioning, not signed. LNG-primary pipelines don’t count unless the operator explicitly names DC load.
How volumes got normalized
Everything shown in MMcf/d. MMBtu/d converts at 1,032 Btu/cf throughout. Beyond that, three different things end up in the same column and you should know which is which:
- Operator-disclosed gas supply. Where the operator’s press release named a specific MMcf/d figure (Homer City 645, Project Jupiter 388, Trident 2,000, EQT-Duke 800), that number is used. Most rows in the top ten rest on this.
- Modeled from MW where MW is what got disclosed. Several large rows describe MW power capacity in primary reporting (Chevron-Microsoft 2,500 MW, VoltaGrid-Oracle 2,300 MW) but no MMcf/d figure. Earlier tracker versions carried these as MMcf/d — a unit error retracted in v2.4.0. They now use modeled steady-state burn at heat-rate × capacity factor (~85% CF for these AI-baseload rows). Each affected row’s
volume_basisfield shows the math. - Modeled from MW. Where only nameplate is disclosed (Dominion Chesterfield, NIPSCO-AWS, Entergy-AWS, Meta-Hyperion, Vistra Permian, NRG Cedar Bayou, xAI Memphis, etc.), MW converts at heat-rate × capacity factor. Canonical default: 75% CF at 7.0 MMBtu/MWh — ~0.122 MMcf/d per MW, the P50 case. Explicit AI-baseload exception: 85% CF for the small set of rows with hyperscaler-direct, behind-the-meter, sustained training/inference loads (Chevron-Microsoft, VoltaGrid-Oracle, xAI Memphis) — ~0.138 MMcf/d per MW, flagged in
volume_basis. P10 (60% CF, 7.5 HR) and P90 (90% CF, 6.5 HR) sensitivities live on each row.
That mismatch is real and worth flagging. The 21.11 Bcf/d gross is the sum of four different volume meanings and should not be treated as "gas physically required Monday." The v2.5.0 schema makes this explicit per row: the volume_derivation field declares whether each row is disclosed (operator stated MMcf/d), modeled_from_mw (gas backed out from nameplate), positioning_only, or watchlist_no_gas_contract. The reconciliation block above shows the four-way decomposition.
The confidence tags
Every claim on this page is tagged. Measured means it’s in a primary source with a specific figure. Reported means named trade press has it, but we haven’t matched it to a filing. Modeled is a number we derived from a disclosed input — MW × heat rate × utilization, that kind of thing. Scenario is forward-looking, with the assumption stated. Observation is a pattern read across the dataset, not a measurement. Every tag has a hover-reveal showing the math.
What we’re not claiming
Not a forecast. Not a call on which deals close, or who takes molecules from whom. The Permian-vs-Appalachia comparison is about signed commitments and announced positioning through Apr 29, 2026 — not total potential and not in-service volumes. Several large Permian producers (EOG, ConocoPhillips, Devon, Marathon, Oxy) haven’t disclosed DC-linked supply deals and aren’t in the totals. The moment one of them does, this page updates and the gap starts closing.
Commitments are not flowing volumes. The 21.11 Bcf/d gross figure aggregates everything from in-service contracts (Entergy Mississippi has units running today) to permitted-but-not-built (Homer City has its turbine order placed; first power expected 2027) to FERC-pending filings (Project Jupiter is on file, not in service). First-gas dates span 2024 through 2030+. The v2.5.0 schema includes a first_power_year field per row (53 of 56 rows have a specific first-power year; 3 are positioning rows with no committed date).
All 56 counted rows have multiple attached sources after the v2.5.0 source-verification sweep — single-source rows reduced from 10 to 0. 54 of 56 rows have at least one primary source (corporate press release, regulatory filing, or SEC filing); only 2 rows currently rest on trade press only and are pending primary confirmation. The "Cites" column shows the source count, the "Evidence" column shows the evidence_class field (A through E). Counted-only distribution: 15 Class A (operator-disclosed gas + primary source), 27 Class B (MW disclosed, gas modeled at heat-rate), 1 Class C (regulatory or trade-press attribution), 4 Class D (multi-trade-press only), 9 Class E (positioning/watchlist/exclusivity/AIP) — total 56. Five additional watchlist rows (Crusoe Goodnight, Solaris >600 MW, etc.) sit outside the gross.
Reproducibility
The whole dataset ships as deals.json. The authoring pipeline is XLSX → Python → JSON. If you find an error, email hello@sunyascoop.com. Every correction gets logged.
Go deeper
Open a chat with Claude or ChatGPT already pointed at this page. Ask it to pressure-test the thesis, flag the weakest deals, or compare the tracker against a portfolio you care about.
Every correction, in public.
Every release ships a changelog. Errors get retracted in public, with the math. The most recent release is on top; everything that preceded it is below.
v3.7.3 — Reviewer fix-now audit: gross/net language, evidence rollup, CF rule, VoltaGrid date, Solaris watchlist, Entergy $57B
- Gross / net language tightened. "Gross gas-related exposure" → "Gross disclosed exposure" (21.11 Bcf/d). "Net" reframed as "named-overlap net" — gross minus the 1.45 Bcf/d named-chain overlap = ~19.66 Bcf/d. Probability-weighted: ~9.14 Bcf/d. Signed-gas-supply only: ~4.89 Bcf/d. The methodology block now leads with that four-line reconciliation so a Goldman/PE reader can't call the headline inflated.
- Capacity-factor rule made canonical. Default 75% CF at 7.0 MMBtu/MWh (~0.122 MMcf/d per MW, P50). Explicit AI-baseload exception 85% CF for hyperscaler-direct, behind-the-meter, sustained training/inference loads (Chevron-Microsoft, VoltaGrid-Oracle, xAI Memphis). Earlier prose carried both 50% CF and 85% CF assumptions in different places — that contradiction is gone.
- Chevron + VoltaGrid combined math corrected. 2,500 + 2,300 MW = 4,800 MW. At canonical 75% CF: ~586 MMcf/d combined. At 85% AI-baseload exception: ~664 MMcf/d. Earlier "~640 MMcf/d" prose was a mid-bucket estimate — replaced everywhere with the explicit two-CF reconciliation.
- VoltaGrid–Oracle date corrected. Original VoltaGrid 1+ GW Oracle deal announced Oct 15, 2025 (Reuters), not April 2026. The Apr 27, 2026 reclassification was the Energy Transfer Project Jupiter green-chile chain, not the original deal.
date_correction_noteadded at row level. - Evidence-class rollup recomputed. Counted-only: 15 A · 27 B · 1 C · 4 D · 9 E = 56. Old prose ("16 A, 23 B, 10 E, 1 D = 50") was stale.
- Solaris >600 MW watchlist row added. SEI flagged >600 MW behind-the-meter mobile/modular gas-fired capacity tied to data-center customers in Q1 2026 results.
excluded_from_total: true; modeled MMcf/d shown for scale only. - Entergy capex updated to $57B. Apr 29, 2026 Reuters: Entergy raised its 4-year capex plan to $57B citing strong DC demand and up to ~12 GW future pipeline. Added to
sponsor_capex_guidanceand inserted at top ofcomparable_transactions. - Language hygiene. "Net gas demand" → "net disclosed exposure" (commitments are not flowing volumes). Grid-operator "takes a cut" replaced with "the utility and grid sit between them as the regulated delivery and reliability wrapper — recovering cost of service through rate base, not skimming a spread." "Skip pure grid procurement" expanded inline: co-located, dedicated, behind-the-meter, or hyperscaler-underwritten. "Seventeen months. Fifty deals." → "Sixteen months. Fifty-six counted deals."
v3.4.0 — Underwriting layer · v3.7.0 — Mobile overhaul
- Power-to-compute maturity score (0–10) per row. Every deal now scored on a deterministic rubric: 0–2 rumor/announced, 3–4 gas source identified / AIP, 5–6 permits filed → approved, 7–8 equipment ordered / FID, 9–10 under construction / energized. Surfaces in the ledger as a colored chip and in the dossier as a labeled progress bar.
- Executive Buyer Mode toggle. New button in the analyst-mode bar re-sorts the ledger by maturity score (deliverable AI MW × execution risk) instead of gross Bcf/d. Same dataset; underwriting-first navigation. Aimed at hyperscaler infrastructure teams, energy underwriters, and credit analysts.
- Two new fields per row.
first_power_quarter(derived from year + binding status) andultimate_ai_buyer(extracted from counterparty + headline + structure — Microsoft, Meta, Oracle, AWS, Google, OpenAI, xAI, Anthropic, CoreWeave, Crusoe, Stargate, etc.). - Reviewer-driven data corrections. Coterra/CPV Basin Ranch corrected from 168 → 50 MMcf/d signed (full plant burn preserved separately). Project Jupiter + VoltaGrid-Oracle reclassified to
volume_class: fuel_cell_burnwith SOFC math (~0.1264 MMcf/d per MW vs turbine's 0.122). Top-ten concentration corrected from 82% → 78% (16.5 ÷ 21.11). Heat-rate explainer math fixed (was 0.165 MMcf/d/MW, now correctly 0.122). Btu/cf inconsistency standardized to 1,032. Waha 2 Bcf/d flaring claim rewritten as "stranded, constrained, or discounted Permian associated gas" with caveats. - Row-level location & status fixes. Project Jade corrected from Cheyenne County, NE → Cheyenne, WY. Nscale Monarch corrected from Texas → West Virginia; Microsoft involvement downgraded from definitive → AIP. SoftBank/AEP/Portsmouth surfaced as 1.12 Bcf/d modeled potential (was hidden as zero). Pacifico GW Ranch full 7.65 GW permit potential surfaced separately from Phase 1 count. Two new watchlist rows: O'Leary/Stratos full-build (9 GW potential), NextEra Japan-backed (~10 GW PA + TX pipeline). xAI Memphis Colossus tagged with permit_litigation + community_opposition risk flags after the Apr 2026 NAACP lawsuit. Fermi/Project Matador split into structured capacity buckets (signed/secured/permit/tenant) with probability cap at 0.40.
- Source-class consistency. 194 source records reconciled —
typefield derived fromsource_classper canonical mapping (COMPANY_FILING/COMPANY_IR → primary; TRADE_PRESS/ANALYSIS/SECONDARY → secondary). Eliminates the "S&P primary + TRADE_PRESS" contradictions. - Mobile overhaul (v3.7.0). Tap targets ≥36px on every interactive element. Hero typography clamped to viewport. Hero stats stack 1-col with hairline rule. Counter-thesis cards stack. Skeptic toggles 1-col. Ledger gets horizontal scroll with sticky first column. Dossier opens full-screen on phone. Floating ⌘K trigger button bottom-right. 3D hex map flattens to 2D. Defensive overflow clipping eliminates accidental horizontal scroll.
- Aesthetic polish. Sparklines under each hero stat showing 16-month cumulative growth. Slope chart "July 2025 — EQT-Southern signed. Appalachia crossed the Permian and never came back." annotation overlay. Operator ticker chips (EQT, WMB, ET, KMI, CVX...) overlay on bar chart. Killed the OS
?hover cursor (was on everycursor: helptag chip). 70 straight apostrophes converted to curly. Animation safety net forces stuck.revealelements to final state after 2s. Focus rings sharpened. Reveal-completion guaranteed on mobile. - Counts & headline. 61 rows · 56 counted · 5 watchlist · 21.11 Bcf/d gross disclosed exposure · 19.66 Bcf/d named-overlap net · 19.6% of US 2026 dry gas production · 0.66 Bcf/d in fuel-cell sub-tier (Jupiter + VoltaGrid).
v3.3.0 — Eight new deals, Jupiter fuel-cell correction, single-canonical-RELEASE drift sweep
- Eight new disclosed deals added. Tier 1 sweep: Certarus CNG-to-hyperscaler, O’Leary/Stratos Utah Phase 1, Hope Gas–Nscale Monarch (WV), Pacifico GW Ranch Phase 1, B&W/Base Electron–Applied Digital, ElectriGen 1.8 GW Texas BTM, ECP-KKR–Calpine Thad Hill (TX). Plus SoftBank/AEP Portsmouth (OH) added to watchlist with
excluded_from_total: true— a real intent signal but no signed gas contract yet. - Project Jupiter fuel-cell correction. Energy Transfer’s Jupiter (Apr 27) and VoltaGrid–Oracle now footnoted: Bloom Energy fuel cells will replace the originally-planned gas turbines for the first phase. The gas commitment remains, but downstream burn assumptions changed.
probability_weighton both rows reduced ~55%;jupiter_bloom_fuel_cell_correctionrisk flag added to deal records. - Single canonical RELEASE. Every visible release/version/count/date now hydrates from the
RELEASEobject and fromcomputeRelease()'s data-derived fields. Title, OG description, hero header, footer, BibTeX, print cover, citation block, Cmd+K palette labels, ledger header — all wired. Eliminates the v3.0 / v3.1 / 49 / 50 / 169 / 153 ambiguity that had been showing up in different parts of the page. - Counts updated. 61 rows · 56 counted · 5 watchlist · 21.11 Bcf/d gross · 197 source entries across 178 unique URLs · 19.8% of US 2026 dry gas production (EIA STEO).
v3.1.0 — Capacity-factor standardization + behind-the-meter cleanup + reader CTA
- Standardized 85% capacity factor for AI-baseload modeled rows. Previous releases used a mixed default — 50% CF for utility/IPP rows (a generic CCGT proxy) and 85% CF for AI-baseload rows (Chevron-Microsoft, VoltaGrid-Oracle, xAI Memphis). The mixed default undercounted gas burn for rows like PPL-Blackstone, Hyperion, NIPSCO-AWS that are explicitly built for sustained AI training/inference loads. v2.6 normalizes 17 modeled rows to 85% CF using EIA’s 2024 CCGT heat rate (6.3 MMBtu/MWh for combined-cycle, 7.2 for reciprocating engines). Each affected row’s
volume_basisfield shows the math. - Behind-the-meter classifier rebuilt. The v2.5 BTM heuristic was conservative and missed Homer City, Shippingport, Hyperion, NIPSCO-AWS, and others that are clearly co-located with their hyperscaler customers. Reset and applied a manual override for 19 rows. Total BTM count: 18 of 50 rows; 7 of the top-10 by MW.
- Net effect on numbers. Gross 18.3 → 20.1 Bcf/d. Net 16.8 → 17.85 Bcf/d. Appalachia 4.6 → 5.1 Bcf/d (PPL-Blackstone and NIPSCO-AWS rerated). Permian 1.5 → 1.9 Bcf/d. App still leads Permian, now by 2.6× (was 3.2x). EQT share of App+Perm: 43% → 38% (denominator grew). ~17% → ~18% of US dry gas.
- Williams "1.4 GW" claim corrected to 2.1 GW (Power Innovation projects sum to 2,140 MW). Williams ~492 MMcf/d producer-equivalent (the "drops to ~630 MMcf/d" figure in earlier releases was an arithmetic error).
- Project Jupiter 380 → 388 MMcf/d per actual ledger value.
- Finding 04 rewritten. Was: "Six of the ten biggest gas sites skip the grid entirely." Actual count of behind-the-meter rows in top-10 by MW was 3 in v2.5 data due to incomplete BTM tagging. After v2.6 cleanup, the count is 7. Finding now reads "Seven of the ten biggest gas-fired AI sites skip pure grid procurement."
- Changelog moved out of the hero flow. Previously sat between hero and argument body — broke first-time reader experience by leading with version-history before any analysis. Now lives as its own section between Methodology and Cite-This. Returning readers can scroll; first-time readers get the argument first.
- Methodology section retracted three fabricated v2.5 changelog claims (NIPSCO-AWS source verification details, ExxonMobil-Calpine "Quail Run → Baytown" fix, Homer City utilization rewording) — these had been described as completed work but weren’t actually done in code. Removed.
- Reader CTA added. Replaces the generic "Subscribe" call-to-action with a higher-friction prompt: "Working on a deal in this space? Tell us about it in one sentence and we’ll send the next tracker before it ships." Aligned with what engaged readers (bankers, PE principals) actually want.
- Schema redesign. Every row now carries 11 new fields:
volume_disclosed_by_source,volume_derivation,mw_disclosed,basin_group,technology,behind_the_meter,ultimate_power_buyer,contract_binding_status,evidence_class(A–E),excluded_from_total, andrisk_flags. Earlier versions mashed disclosed gas, modeled-from-MW, pipeline capacity, and positioning rows into one column; now each row declares which kind of volume it represents. - Diamondback Permian JV retracted to positioning. Bloomberg, NGI, DCD, and World Oil all describe Diamondback as "in talks" for an exploratory power JV — no MW disclosed, no gas volume disclosed, no counterparty. Previous tracker versions carried this row at 600 MMcf/d / 2,000 MW, which had no source backing. Corrected to
mmcfd: 0with status = Positioning. - xAI Memphis expanded to capture full scope. Previously listed at 460 MW (Memphis only), now 1,450 MW combined: 247 MW permitted Memphis + ~1,200 MW Southaven, MS (41 newly permitted turbines March 2026 + 27 unpermitted turbines currently running). NAACP/Earthjustice/SELC filed federal Clean Air Act lawsuit April 2026 over the unpermitted turbines. Modeled at 206 MMcf/d.
- Crusoe Goodnight added as watchlist row. 933 MW gas plant permit filed at Crusoe’s Goodnight, TX campus per WIRED reporting; Google involved as customer. No gas operator counterparty disclosed, so excluded from the 18.3 Bcf/d gross total via
excluded_from_total: true. - Source-verification sweep on top 10 rows + single-source rows. Web-searched primary sources for the top 10 rows by volume. Diamondback Permian JV retracted per the verification — no source backed the 600 MMcf/d / 2,000 MW figures. Range Resources and ET Desert Southwest received added primary sources. Single-source rows reduced from 10 to 0; no-primary rows reduced from 9 to 2.
- Evidence-class distribution. 16 rows Class A (operator-disclosed gas + primary source), 23 rows Class B (MW disclosed, gas modeled), 1 row Class D (multi-trade-press only), 10 rows Class E (positioning/exclusivity/AIP/watchlist).
- Net effect on numbers. Gross 17.9 → 18.3 Bcf/d. Net 16.4 → 16.8 Bcf/d. Permian producer supply 2.1 → 1.94 Bcf/d (Diamondback retraction). Appalachia 4.6 Bcf/d (unchanged). App leads Permian by 3.1 Bcf/d, more than 3×. EQT share of App+Perm: 40% → 43%. Top 10 share: 83% → 84%. ~16% → ~17% of US dry gas (DSW upsize).
- Chevron-MSFT and VoltaGrid-Oracle were unit errors. 2,500 MW and 2,300 MW carried as MMcf/d. Bloomberg, Yahoo Finance, DCD, Power Magazine, and VoltaGrid’s own press release all describe MW power capacity, not gas volume. Both rows corrected to modeled steady-state burn (Chevron 311 MMcf/d, VoltaGrid 327 MMcf/d).
- The headline thesis flipped. "Permian caught up" was wrong because both major Permian rows were inflated. Appalachia leads Permian by more than 2× on disclosed producer supply.
- Net effect. Gross 22.0 → 17.9 Bcf/d. Net 20.5 → 16.4 Bcf/d. EQT share of App+Perm: 25% → 40%.
- Producer-supply numbers refreshed. Appalachia 4.3 → 4.6 Bcf/d. Permian 5.9 → 6.1 Bcf/d (later corrected to 2.1 in v2.4.0). Deal counts: Appalachia twelve → fifteen, Permian eleven → thirteen.
- Volume-basis transparency. All 56 rows received
volume_basisfields documenting how each MMcf/d figure was derived.
- Four arithmetic errors fixed. Williams Socrates 120 → 50 MMcf/d. Williams Socrates Younger 110 → 42. Three MMBtu/d ↔ MMcf/d conversions normalized to 1,032 Btu/cf.
- Volume-basis distinction made explicit. Chevron-MSFT (2,500) and VoltaGrid-Oracle (2,200) come from contract-envelope reporting, not from heat-rate calculations.
- "Ten names carry ninety percent" was wrong. Top ten = 86%. Corrected throughout.
- Mid-Atlantic + AWS + xAI + IPPs + Meta right-sized. Dominion Chesterfield (944 MW), NIPSCO–AWS Indiana (2.6 GW), Entergy–AWS Warren County (754 MW), xAI Memphis Colossus (460 MW), Meta-Entergy Hyperion expanded to 7+ GW, Vistra Permian (860 MW), NRG Cedar Bayou (1.5 GW), Vistra-Cogentrix positioning (5.5 GW). Closed AWS, Mid-Atlantic, IPP, and xAI gaps.
Know a deal we missed?
Send one sentence. Anonymous tips welcome. Corrections are published openly with the math. Source links are credited if you want — held confidential if you don’t. Sensitive deal flow can be discussed off-list. Replies read by Raj. No mailing list. No automation. No autoresponder.
The bottlenecks behind the bottlenecks.
Five layers of public data that frame the per-row deals: turbine OEM backlogs (where equipment scarcity actually binds), ISO interconnection queues (where the grid bottleneck lives), basin basis history (where pricing creates the spread), hyperscaler capex (where the demand sits), and the NREL benchmark LCOE (where the investability spread is measured against). Underwriting starts here.
Buyer rollups are Sunya-attributed based on named counterparties, tenants, or explicit project association. They are not self-disclosed aggregate Bcf/d commitments by Microsoft, Amazon, Meta, Oracle, or Google. The 51% figure is a count of disclosed deals where the named hyperscaler is the economic buyer or campus tenant — not an admission by either company.
Public reference for evaluating per-row LCOE estimates. Project-specific estimates in our dataset are computed via the formula in each row's lcoe_basis_note; this is the sanity-check baseline.
Recent prints in gas-to-AI infrastructure.
Seven public-record transactions in the last 12 months that template what's investable. Includes Ares/Rover (the midstream-stake-from-Blackstone deal that defined "strategically located gas egress"), Applied Digital's $7.5B 15-year hyperscaler lease (showing single-tenant lease structures support project finance), GE Vernova's 110+ GW backlog disclosure (proving equipment scarcity is the binding constraint), Entergy/Meta's revised structure (the ratepayer-protection template), and Project Jupiter's fuel-cell pivot (showing technology substitution risk).
Counterparty concentration
Top buyers across the counted deals. Investment-grade hyperscaler exposure dominates.
Public sponsor capex guidance
2026 capex disclosed by named operators on most recent earnings or investor-day calls. Approximate data-center allocation as a % of total capex.
Three scenarios for 2028.
What actually delivers? The base case projects the current pipeline materializing per maturity scores. Bear case haircuts for AI capex slowdown + equipment delays + permit fragility. Bull case assumes equipment slots free up + capacity-market inflection economics + pre-FID rows accelerate. Wide spread reflects high execution-risk uncertainty.
How hyperscalers actually buy.
Each named hyperscaler / AI buyer scored on speed-to-power, ESG floor, contract flexibility, ratepayer-protection requirements, and disclosure transparency. Same dataset; different vendor decisions. Microsoft is most disciplined; Oracle is fastest; xAI is fastest-and-most-fragile; Meta is most community-protective.
The macro permitting backdrop.
Trump administration executive orders + 2025-2026 regulatory action affecting permitting speed, NEPA review, IRA implementation, and EPA emission thresholds. Each shifts the baseline for individual project timelines. EO 14156 (energy emergency) is the most consequential — it invokes emergency authorities to expedite gas pipelines, plants, and transmission.
Every retraction since v2.0.
Every correction we've made since the first reconciled edition. Each entry shows the version, the date, the specific correction, and the net Bcf/d impact. Transparency is the operating principle.
What every field means, what's verified, what isn't.
Every per-row field we add comes with a confidence label. Verified = public-record-confirmed fact. Inferred = derived from existing fields via deterministic rule. Estimate = formula-derived range. Low confidence = honest default when no public source. Not disclosed = explicitly marked as unknown.
Use it. Cite it. Break it.
The whole dataset is open. Grab the JSON, export the ledger as CSV, copy a clean citation. If you publish something that builds on this — critical or otherwise — email a link and we’ll link back.
How to cite
Mistry, Raj. The Gas-to-AI Tracker: The AI gas boom looks Texan in the headlines. It looks Appalachian in the contracts. Sunya Research, Reconciled Edition v3.8.5, Apr 29, 2026. gastoai.sunya.ai
AI's gas boom is not being misread because people missed the demand. They counted the wrong thing. The loudest announcements are measured in megawatts; the binding advantage is measured in molecules. Once the April Permian megadeals are converted from power capacity into modeled gas burn, Appalachia still leads on producer-attributed exposure, EQT remains the dominant named supplier, and the prize is clear: control the chain from gas origin to pipe to turbine to utility wrapper to hyperscaler credit. Not a data-center siting story. The birth of the gas-to-compute infrastructure stack.