Energy Supply Strategy: Grid PPA, BYOP & Co-Location

By 2026 the binding question is no longer "can I buy the chips?" but "can I energize them, and at what price, firmness, and carbon profile?" Chapter 3.2 established that the grid queue is the long pole — 3 to 7-plus years end-to-end, over 8 years application-to-COD in PJM — and Chapter 3.3 established that power is 25–60% of the cost stack and the dominant TCO lever. This chapter is where those two facts collide into a single strategic artifact: the energy supply portfolio. It is the decision about where your electrons come from, on what contract, at what price certainty, and against what regulatory and market risk — and it is one of the least reversible commitments in the project, because the contract tenor, the interconnection class, and the on-site generation all outlive the GPU generation they were scoped to serve.

The chapter works through the supply portfolio in layers. We frame the four supply structures and the fork between them; we take apart PPA structures — physical versus virtual, fixed versus indexed, and the 24/7 carbon-free-energy matching debate that quietly drives firming cost; we treat power-market risk management as a first-class engineering subject (nodal basis, VPPA shape and negative-settlement risk, congestion-revenue rights, heat-rate hedging) because an ERCOT-merchant exposure can swing a pro-forma further than any cooling optimization; we lay out long-duration firm supply (nuclear restarts, uprates, SMRs, renewables-plus-storage firming); we work the co-location regulatory treatment that the December 2025 FERC PJM order reshaped; and we close on the tenor-matching problem that ties a 17-year contract to a 2–3 year chip. The through-line: most of these structures trade one risk for another — they do not eliminate risk, they relocate it, and the operator's job is to know which risk it has chosen to hold.

Every energy supply strategy is a blend of four archetypes, and naming the dominant one is the first decision. Grid-only takes firm utility power through a standard large-load interconnection: lowest unit cost at scale, full reliability obligation on the utility, and the slowest path — you wait in the queue, and the queue is the constraint. Grid + bridge generation energizes early on temporary or permanent on-site generation (typically gas turbines or reciprocating engines) and transitions to grid power as the interconnection completes: it buys 18–36 months of speed-to-power at the cost of fuel, emissions permitting, and the risk that the "bridge" becomes permanent. Behind-the-meter (BTM) island bypasses the grid entirely with dedicated on-site generation: maximum speed and independence from the queue, but you now own the reliability problem, the fuel-supply problem, and the firmness problem yourself. Hybrid / co-located sites the load adjacent to a generator (gas, nuclear, or renewable-plus-storage) and draws primary power from it while contracting some grid service for backup or supplemental supply — the structure that the FERC PJM order of December 2025 finally gave a regulatory home.

The reason this is a portfolio rather than a single pick is that the four structures hedge different failure modes, and the 2026 default is to layer them: a co-located gas plant for first power, flexible grid service underneath it, a long-dated clean PPA for the carbon claim, and storage for ride-through. The behind-the-meter wave makes the point — roughly 82 GW of BTM data-center generation has been announced since January 2025, yet only ~2 GW was online by mid-2026 (Cleanview / SemiAnalysis, 2026). The announcements are a portfolio bet on speed; the thin operating base is the reminder that bridge power is a schedule instrument, not a finished plant. → on-site generation engineering in Chapter 3.5.

Whatever the physical supply structure, most large operators layer a power purchase agreement on top to fix cost and substantiate a clean-energy claim — and the PPA's structure determines whether you have actually hedged anything. The first fork is physical versus virtual. A physical PPA delivers electrons (or their grid equivalent) to your node: you take title to the energy and the renewable attributes, and your power cost is genuinely fixed at the contract price for that load. A virtual PPA (VPPA) is a financial contract-for-differences: you do not take the power, you settle the difference between a fixed strike and the floating market price at the project's node, and you separately buy your actual power from the grid. The VPPA lets a buyer support a renewable project anywhere — including a different grid from the one your data center sits on — and claim the renewable energy certificates, which is why it dominates corporate clean-energy procurement. But it introduces a risk that a physical PPA does not: the hedge settles at the project's node, while you pay for power at your node, and those two prices diverge.

The second fork is fixed versus indexed. A fixed-price PPA locks $/MWh for the term — maximum cost certainty, but you forgo any benefit if market prices fall and you over-pay relative to a cheaper future. An indexed (or collared, or hub-settled) PPA floats with a market reference inside a band — less certainty, but it tracks the actual cost of power and avoids the worst over-pay outcomes. For a power-bound AI build the instinct is to fix everything, because Chapter 3.3 showed power is the dominant opex line; but fixing the price at the wrong node, on a project whose generation profile does not match your 24/7 load, can create a hedge that loses money exactly when you need it. A PPA relocates risk rather than reducing it, and a badly-shaped one relocates risk toward you.

The carbon claim attached to a PPA is itself a fork, and it drives real cost. The legacy standard is annual REC-matching: buy enough renewable certificates over a year to cover annual consumption, regardless of when the clean power was actually generated. It is cheap and it is increasingly seen as accounting fiction, because a data center running 24/7 on a grid that is dirty at 2 a.m. is not actually carbon-free at 2 a.m. no matter how many midday solar RECs it holds. The rigorous standard is 24/7 carbon-free energy (CFE): matching every hour of consumption with carbon-free generation on the same grid, measured hourly. Google's CFE program is the reference implementation, and its numbers show how hard the last hours are — a global average CFE score of ~66% in 2024, with 9 of 20 grid regions above 80% but Asia-Pacific near 12% (Google 2025 Environmental Report). The gap between 66% and 100% is the firming problem: the cheap renewable hours are easy, and closing the dark, windless hours requires storage, firm clean supply, or over-build — all of which cost money the annual-REC buyer never pays.

The related concept is additionality: whether your procurement caused new clean generation to be built, or merely bought certificates from a project that would have existed anyway. Additional, hourly-matched supply is the defensible claim; non-additional annual RECs are the one regulators and customers are learning to discount. The consequence is direct: commit to 24/7 CFE and you have implicitly committed to a firming portfolio (storage, nuclear, or geographically-diverse renewables) and possibly to a site with good clean-resource availability — a carbon target reaching back into site selection. → operational stewardship in Chapter 15.4; grid-services framing in Chapter 15.8.

An operator with merchant exposure — anyone drawing grid power without a fully-firming physical hedge — is running a power-trading book whether it knows it or not, and the risks are specific and quantifiable. Four matter most.

Nodal / locational basis risk. Wholesale power clears at locational marginal prices (LMPs) that differ node-to-node because of congestion. If your VPPA settles at a renewable project's node in West Texas and your load sits at a node near Dallas, the two LMPs diverge whenever the transmission between them congests — and the divergence is the basis, an uncovered exposure even though you are nominally hedged. In ERCOT, where the merchant market is the default and renewable generation is concentrated far from load, basis can dwarf the energy price itself. Congestion-revenue rights (CRRs / FTRs) are the instrument that hedges basis: a financial right that pays the congestion difference between two nodes, letting you reconnect your generation node to your load node. Buying CRRs to cover a VPPA's basis is the difference between a hedge that works and one that fails at the worst moment.

VPPA shape / settlement risk. Beyond basis, the VPPA's generation profile rarely matches your load profile. A solar VPPA generates at midday and settles against you (the negative-settlement trap above); a 24/7 load consumes flat. The mismatch — the shape — is an uncovered position even when the basis is hedged. Storage-pairing, hybrid wind-plus-solar portfolios, and explicit shape modeling are the mitigants.

Heat-rate / spark-spread risk. For an operator running on-site gas (BTM or bridge), the cost of power is a function of the gas price and the plant's heat rate. A heat-rate hedge (or tolling structure) fixes the conversion of fuel cost to power cost, insulating the budget from a gas-price spike — relevant because the BTM gas wave has tied a large slice of new AI capacity directly to natural-gas markets. Industrial gas-fired LCOE already runs above ~$120/MWh, above projected ERCOT grid prices, so the on-site operator is long power-price volatility and short any hedge it does not deliberately buy.

The synthesis: a PPA without a basis hedge, a shape hedge, and (for gas) a heat-rate hedge is a partial hedge that can lose money in the tail. The honest posture is to map every leg of the supply portfolio to the risk it covers and the risk it leaves open, and to price the residual. → the ERCOT-merchant downside is stress-tested in Chapter 1.8; lender downside in Chapter 2.5.

The carbon target and the firmness requirement meet in the search for firm clean supply — power that is both carbon-free and available every hour. Three paths, on very different timelines. Existing-plant nuclear restarts and uprates are the only firm clean supply deliverable this decade: the Three Mile Island / Crane restart (835 MW for Microsoft, ~$1.6B, targeting ~2028) and the Amazon-Talen ~2 GW arrangement are the templates, and more than 10 GW of nuclear has been contracted by hyperscalers. They work because the plant already exists and is already interconnected — the speed comes from reusing a finished asset, not from anything fast about nuclear. Small modular reactors (SMRs) — Kairos (~500 MW for Google), X-energy, Oklo — are the structural 2030s answer, not a near-term speed fix: first-of-a-kind capex runs $6,400–12,700/kW and realistic delivery sits in 2030–2035, gated by NRC licensing. Renewables-plus-storage firming is the third path: over-build wind and solar, add batteries (and increasingly long-duration storage), and engineer toward a 24/7-CFE profile without nuclear — the cheapest clean MWh but the hardest to make firm in the dark, windless hours, which is exactly where the over-build and storage cost lives.

The choice is a timeline bet. If you underwrite nuclear or SMR firm supply as bankable near-term capacity, you may strand a campus waiting for a 2032 reactor; if you treat it as optionality and firm the gap with gas or storage, you carry the bridge cost and the emissions exposure in the meantime. The honest 2026 posture: existing-plant restarts are real and bankable where you can secure one; SMRs are optionality, not a schedule input; and renewables-plus-storage is real but its firmness is only as good as its over-build ratio and its storage duration. → on-site and nuclear engineering depth in Chapter 3.5.

Co-location — siting the load directly with a generator and drawing power on the generator side of the interconnection point — is the structure that promises the most (queue-bypass speed plus firm clean power if the generator is nuclear) and that lived in the deepest regulatory grey zone until late 2025. The unresolved questions were genuine: if a data center draws power that never touches the public grid, should it pay for the transmission system it occasionally leans on? Who carries the cost when the co-located load needs grid backup during a generator outage? And does pulling a large nuclear plant off the grid to serve one customer harm the ratepayers who relied on that capacity for reliability and price?

The FERC PJM order of December 18, 2025 answered the first cut of these. Finding PJM's existing tariff "unjust and unreasonable" for lacking terms for generators serving co-located load, FERC directed PJM to create new contract-demand transmission services — firm, non-firm, and interim non-firm — so a co-located load can contract for a defined slice of grid capacity rather than free-ride or be excluded, with a compliance window opening in early 2026 and a multi-year transition (to ~2028). The order also reformed behind-the-meter generation cost-causation to shield ratepayers and signaled that Capacity Interconnection Rights (CIRs) — the generator's right to inject capacity into the grid — must be adjusted when that capacity is redirected to a co-located load. The order is PJM-specific but is treated as a template for MISO, SPP, and ERCOT. The consequence for a co-location strategy: the regulatory path is no longer a grey zone, but it is now a priced path — you contract for the grid service you use, you face cost-causation charges that earlier deals avoided, and the CIR treatment can change the generator economics that underpinned your offtake. The bypass got legal; it also got a bill. → grid-impact process mechanics in Chapter 3.2; grid-services revenue in Chapter 15.8.

The deepest structural tension in energy supply is a duration mismatch. The firm clean deals that anchor a campus run 15–20 years — Microsoft-Constellation at 20 years, Amazon-Talen at 17 — and a behind-the-meter plant or a co-location interconnection is a 20–30 year physical asset. But the GPU fleet those electrons power has a 2–3 year economic life (the contested figure that Chapter 1.8 turns on). You are committing to power for a fifth of a century to serve a chip that is obsolete in three years and a workload that might move in five. The mismatch is not academic: a take-or-pay power contract or a co-location agreement does not care whether the GPUs sitting behind it are still earning revenue. If the workload migrates, the demand softens, or the residual market for the hardware collapses, the power commitment remains — a fixed liability against a vanished revenue stream.

The mitigants are all forms of optionality. Match the firm tenor to the durable part of the load, not the speculative part — contract long-dated power only for capacity you are confident will run for the term, and meet the uncertain incremental demand with shorter, more flexible supply (grid, spot, or curtailable). Keep the supply portfolio re-mixable: a co-location anchor plus flexible grid service plus a storage layer can shed or shift cost as the workload changes, where a single 20-year fixed-price BTM commitment cannot. And price the power against the economic GPU life, not the contract life — if the energy strategy only pencils when amortized over 20 years of GPU operation, it does not actually pencil, because the GPUs will not be there. The right posture mirrors the reversible-vs-irreversible discipline of Chapter 1.8: the irreversible firm commitment is sized to the workload you are certain of; the reversible flexible supply absorbs the rest.

The recurring energy-supply mistakes all come from treating a risk-relocation instrument as a cost-reducer, or from matching the wrong durations:

• The unhedged-basis VPPA. Signing a virtual PPA at a remote renewable node to "fix" power cost, with no CRRs/FTRs to close the basis to the load node. The hedge settles where the project is; you pay where the load is; the gap is an uncovered position that blows out exactly when congestion is worst. A partial hedge sold internally as a full one.
• The single-technology solar VPPA on a 24/7 load. Buying a solar-only VPPA to power a flat round-the-clock load, then discovering it settles against you every sunny hour (cannibalization) and supplies nothing in the dark hours. The shape mismatch turns the hedge into a structural drain.
• The permanent bridge. Building on-site gas as a 24–36 month bridge to grid power, then watching turbine lead times and queue delays stretch until the "bridge" is the permanent plant — now carrying fuel risk, emissions liability, and stranding risk it was never sized to hold.
• The tenor mismatch. Signing a 17–20 year firm power contract sized to a peak load that depends on a workload with a 3-year economic life, so a workload migration leaves a take-or-pay liability against a campus that no longer earns. → Chapter 1.8.
• The grey-zone co-location pro-forma. Underwriting a co-location deal on pre-FERC-order economics, where the load free-rode on transmission and CIR value — a model the December 2025 ratepayer-shielding rules are now empowered to unwind.