Tax Incentives, Fiscal Structuring & Economic Development

Incentives are where the strategist and the engineer most often talk past each other. The strategist sees a sales-tax exemption worth tens of millions and a property-tax abatement that cuts a recurring operating cost, and ranks the site accordingly. The engineer sees that none of it delivers a megawatt, a permit or a week of schedule, and that the cheapest site net of incentives can still be the one that never energizes. Both are right. Incentives decide which of two buildable sites you pick; they never make an unbuildable site buildable.

Speed-to-power displaced incentives as the #1 screen years ago (Chapter 3.1), because an interconnection slot is the scarce asset and a tax break is not. Once power, water and permitting are settled, though, incentives move the TCO hard. A full sales-tax exemption on a multi-billion-dollar equipment spend and a large property-tax abatement together shift two of the biggest controllable lines in the model. You will model them either way. The open question is how much of the modeled value you are allowed to believe when you underwrite a 20–30 year asset against incentives that are increasingly 10-year and politically contingent.

The incentive package is a stack of distinct instruments. Each hits a different line of the model and carries a different durability. Treating them as one thing is the first error, because cash-flow shape and reversal risk differ by an order of magnitude across the stack.

Sales- and use-tax exemptions are the largest single lever because the taxable base is equipment. Epoch's May 2026 model puts servers at about 56% of upfront capex and networking at about 13% (Chapter 1.8). Only purchases your project's statute covers are exempt, so the saving is that eligible spend times the rate your program actually exempts. Texas's standard program exempts the 6.25% state tax only, with local tax still due, for 10 or 15 years depending on investment. Its Qualifying Large Data Center Project tier ($500M, 250,000 sq ft, 40 jobs) runs 20 years and covers state and local tax. On a multi-billion-dollar spend the benefit is in the hundreds of millions. Replacement hardware stays exempt only while your certification holds and the exemption term runs, so time refreshes against Chapter 1.8's earning-life scenarios. That equipment base is why the cost to states has exploded. The Texas Senate projects about $3.3B of forgone sales tax over 2028–29 (checked 2026-10-12). Georgia's January 2026 report puts the combined state and local cost of two equipment exemptions at $2.47B for FY2026 and $2.99B for FY2027. The dedicated data-center exemption accounts for $1.16B of the FY2026 figure. Virginia's exemption benefit reached about $1.94B in FY2025 (Va. Dept. of Taxation).

Property-tax abatements and PILOTs cut a recurring operating cost by reducing the tax on eligible property for a fixed, negotiated term. In Texas, Chapter 312 lets local governments negotiate abatements of up to 100% of eligible incremental value, lasting at most 10 years. Negotiated energy rates (discounted or specially structured utility tariffs) attack the largest operating-cost category. In Epoch's May 2026 model of a US hyperscaler running 1 GW of IT capacity, power accounts for about 65% of cash opex; on an annualized TCO basis it is 7% (Chapter 1.8). These rates increasingly come bundled with the large-load tariff terms below. Cash grants, infrastructure cost-sharing and training credits are the smallest dollar lines, but the most politically visible and the most conditional.

Federal tax treatment is not a siting incentive, because it applies at every US site, but it moves the same TCO lines. The One Big Beautiful Bill Act (July 2025) made 100% first-year depreciation permanent for qualified property acquired after January 19, 2025 (IRS Notice 2026-11). That covers servers and network gear, though not the building shell. Section 48E credits can also offset part of what you spend on eligible clean generation and on-site storage (Chapter 2.5).

Every incentive comes with conditions, but a statute and a negotiated contract bind you in different ways. Each program sets its own bar on investment, facility size, timing and certification, plus whatever employment terms it imposes. Of the 38 states offering data-center incentives in NCSL's April 2026 inventory, 23 require job creation and at least 11 set wage conditions. Texas's standard program asks for 20 or more qualifying jobs paying 120% of the county average weekly wage. Model your permanent staffing plan against the covenant of the program you choose. Jobs matter because the political defense of an incentive rests on employment, and an AI data center is among the least labor-intensive billion-dollar facilities ever built.

The PILOT (payment in lieu of taxes) is the structuring workhorse that makes a property-tax abatement legible to both sides. Rather than abate the tax to zero, the developer and the host jurisdiction agree a fixed schedule of payments, often escalating, that replaces the volatile assessed-value-times-millage calculation. The developer converts an uncertain, reassessable liability into a fixed, modelable cost. The host converts a politically toxic "they pay nothing" optic into a concrete revenue line. Host-fee and community-benefit terms increasingly attach to the PILOT too (Chapter 3.11).

The clawback turns the incentive from a subsidy into a risk-shifted obligation. It lets the jurisdiction recover some or all of the granted benefit if the developer misses a covenant: falls short on jobs or investment, or misses the in-service date. For underwriting, a negotiated incentive is not a clean cost reduction; it is a reduction paired with a contingent liability, sized by the gap between what you promised and what the workload actually requires. A facility that pivots from a labor-heavy mixed-use design to a lights-out inference hall has enlarged its own clawback exposure without touching the incentive line.

The most consequential fiscal structure of the 2026 era is not a tax incentive. It is the negotiated large-load tariff, the special rate class that utilities and regulators built to make large AI loads pay for themselves instead of shifting cost to residential ratepayers. It belongs in this chapter because the developer trades a fiscal commitment for a grid outcome: you accept a binding cost floor in exchange for the utility's willingness to plan, contract and cost-allocate your service. The tariff sets the terms, while the studies and physical upgrades still set the timeline (Chapter 3.2).

Oregon's POWER Act (HB 3546, 2025) and PGE's Schedule 96 set the template. The Oregon PUC approved it in May 2026 (Order 26-154, UM 2377), and service under it began July 8, 2026. It is now spreading: Pennsylvania's PUC adopted its own model large-load tariff on April 30, 2026 (order entered May 12), and as of September 2026 EEI counted approved large-load tariffs in 25 states, with seven more pending. The template bundles four terms every large-load developer must now price. (1) The customer pays 100% of the cost of distribution facilities that serve it exclusively, and a proportional share of facilities it shares with others (PGE Rule I, effective July 8, 2026). (2) Minimum generation and transmission demand charges at 90% of contracted system capacity, billed whether you use it or not: a take-or-pay floor. (3) Contract tenors of 10–30 years scaled to load size, 30 years at 220 MW and above, with early-termination charges tied to the remaining demand obligation and the unrecovered distribution investment. (4) A 1¢/kWh surcharge at 100 MW or more of allocated system capacity, funding residential cost-offset and low-income programs. The class starts at 20 MW, counting demand, contracted demand and aggregated campus tenants. These are the PUC template figures. Verify the operative Schedule 96 and UM 2377 sheets and your project service agreement before pricing a floor, because thresholds, exemptions and the serving utility's own tariff differ (Pacific Power's UE 463 is a separate proceeding). The result makes cost-causation contractual: the load that causes the upgrade pays for it and commits to use enough of it that the upgrade is not stranded.

California legislated its version on September 21, 2026. SB 886 has the CPUC set data-center tariffs by January 1, 2028, with an early-termination fee for a customer that leaves within 10 years of interconnection or fails to ramp its load. AB 2383 adds ten-year minimum payments, early-termination fees and upfront collateral, and puts wildfire, electrification and other public-purpose charges on the data-center tariff. The CPUC may set the tariff's demand threshold no higher than 25 MW. SB 1168 orders a CPUC review of how transmission and distribution costs are allocated (Mayer Brown, September 28, 2026). Until the CPUC rules, a California site knows the shape of its tariff but not the rate (Chapter 3.3).

The take-or-pay minimum-demand charge converts a utilization risk into a fixed-cost risk. A campus that ramps slower than planned, or whose workload demand softens, still pays 90% of contracted system capacity under the Schedule 96 template. An empty hall still incurs that bill, and fewer billable GPU-hours or a lower rental price leave less contribution to cover it alongside debt service. Carry the tariff floor into Chapter 1.8's matched cash-flow comparison and Chapter 2.5's DSCR case. Under the tariff, a demand miss is a penalty, not just a margin squeeze. Sign the tariff when speed-to-power dominates and your ramp forecast is credible. If demand is uncertain enough that the 90% floor plus the exit charge would bleed cash into an empty hall, consider a flexible or curtailable option where the utility offers one. Price that option first, because it carries its minimum bill, term, security and exit obligations too.

Incentive structuring does not travel across borders. The instrument, the magnitude and the way you lose it all change by jurisdiction, so a global siting program must underwrite three distinct fiscal regimes.

The United States: deep, fragmented and now contracting. The US offers the deepest and most varied incentives: full sales-tax exemptions, negotiated property abatements, PILOTs and negotiated utility rates. They are fragmented by state and county, and as the warning above shows, they are in active rollback. The US durability risk is legislative. Today's incentive is real and large. A state legislature can still sunset or repeal it on a one-to-two-year political clock, and the statute or agreement decides whether existing projects are protected.

The European Union: shallower, slower and state-aid-constrained. EU incentives are bounded by the state-aid regime: the General Block Exemption Regulation (GBER) and notification rules that cap how much public support a member state can grant a single undertaking without Commission approval. There is no US-style blanket sales-tax exemption on equipment. Support tends to flow through energy-cost relief, grants for green infrastructure and regional-development funds. Each needs the operative GBER category, eligible-cost basis, aid intensity and notification threshold, cumulation rules and any Commission clearance. Get the consolidated regulation and the awarding authority's legal basis before booking a grant. The EU is mid-reform: a new GBER was in consultation in early 2026, with adoption targeted for end-2026 and entry into force in January 2027, and the digital sector is explicitly in scope. The EU durability risk is regulatory. A package can be challenged as unlawful state aid and recovered with interest years after the fact, which is a different and in some ways harder risk than a US legislative sunset.

The Nordics show how a European tax break unwinds. Three Nordic countries gave data centers the low electricity-tax rate reserved for heavy industry, and all three have taken it back. Norway's reduced rate ran from 2016 to 2022; the 2023 state budget abolished it, and data centers have paid the ordinary rate since January 1, 2023. Sweden ended its 2017 data-center reduction on July 1, 2023, moving operators from 0.6 öre/kWh to the ordinary service-sector rate, about 39 öre/kWh that year (Riksdag decision, December 2022). Finland moved data centers from electricity-tax class II (0.05 c/kWh) to the general class I (2.24 c/kWh) on July 1, 2026 (government bill HE 156/2025; Finnish Tax Administration). Denmark never had a data-center rate. It cut electricity tax for every user to 0.8 øre/kWh for 2026 and 2027 (Danish Tax Agency). A VAT-registered operator running its own servers could already reclaim the tax on that power down to the EU minimum of 0.4 øre/kWh, so the cut mainly helps a colocation landlord, which cannot reclaim tax on its tenants' server load (Skatterådet binding rulings, 2016 and 2018). Underwrite a Norwegian, Swedish or Finnish site at the ordinary rate, and a Danish colocation site as if the cut ends in 2027. In Norway, the Storting did not adopt a February 2026 motion to study a higher data-center rate (Innst. 183 S, March 26, 2026; debated April 21, 2026).

The Middle East and APAC special zones: sovereign incentives at a different scale. The Gulf substitutes sovereign incentives for the negotiated patchwork: free-zone corporate-tax holidays, subsidized or sovereign-supplied power, state co-investment and land at nominal cost, all wrapped inside national AI strategies (UAE/Stargate; Saudi HUMAIN, which targets roughly 3 GW by 2030 as of October 2026). In the UAE, a Qualifying Free Zone Person may pay 0% corporate tax on qualifying income, as long as it passes economic-substance tests that are getting stricter. Multinational groups in scope are also subject to the domestic minimum top-up tax. From financial years starting January 1, 2025 or later, a group is in scope if its consolidated revenue hit €750M in at least two of the four years before. The top-up lifts the jurisdictional effective rate toward 15% under GloBE rules, so model it with whatever exclusions and relief apply (UAE Ministry of Finance; FTA scope guidance, October 2026). APAC special economic zones (Malaysia's Johor, India's IT/SEZ regimes, Singapore's targeted schemes) offer zone-specific holidays and duty exemptions. Underwrite the tax-law and qualification risk together with the sovereign commitments and the chip export-control conditions covered in Chapter 3.12. Those conditions depend on a sovereign relationship and a US export-control posture that no local legislature controls.

Brazil ties a federal tax suspension to engineering conditions. Its Redata regime (a special tax regime for data-center services) first came by provisional measure, MP 1.318 of September 2025, which lapsed on February 25, 2026 when Congress did not convert it. The same regime returned as an ordinary bill and became Lei 15.504, signed September 15, 2026. It suspends PIS/Cofins, IPI and, for goods with no Brazilian equivalent, import duty on data-center equipment, and the suspension becomes an exemption once the commitments are met. The commitments are design terms. All its electricity must come from renewable or low-emission supply, and water-usage effectiveness must be 0.05 L/kWh or less, which in practice means closed-loop or dry cooling (Chapter 3.7). At least 10% of capacity must serve the Brazilian market, research institutions or the public sector, or the company instead invests an extra 10% of the benefited purchases in research and development. Separately, 2% of the benefited purchases must go into Brazilian research and development (8% and 1.6% in the North, Northeast and Center-West). Miss the energy, water or research terms and you repay the suspended taxes with interest and fines; miss the 10% domestic share and new purchases lose the suspension, with enrollment cancelled if the shortfall is not cured within 180 days of notice (Senate news agency, September 16, 2026). The runway is short: Brazil's tax reform replaces PIS/Cofins with the new CBS on January 1, 2027 and the law does not mention CBS, so unless Congress extends it, little beyond the import-duty relief survives 2026 (Huland, Castro Alves, Linhares & Barros Leal, September 2026).

Everything above has to become a number you can defend in a board package and a lender's model. The core error is treating incentives as a contractual cash flow when most of the stack is a risk-weighted expectation. A signed PILOT with a fixed schedule and no early-termination right is close to contractual. A statutory sales-tax exemption the legislature can repeal, or an EU grant that could be recovered, is not. Give each instrument its own durability haircut and discount it accordingly, rather than dropping the gross incentive value into TCO as if it were guaranteed.

Model it three ways and disclose the spread. (1) Gross case: full incentive value over its quoted term, the number the economic-development office gives you. (2) Risk-weighted case: each instrument multiplied by a survival probability over the asset life, with a steeper curve for statutory exemptions (legislative repeal risk) than for signed PILOTs (contractual, but clawback-exposed). (3) Stress case: incentives sunset or are repealed at the next plausible legislative inflection, realistically the 2027–2029 sessions for a US build started in 2026, with grandfathering assumed only where the statute or agreement explicitly provides it. The gap between the gross and stress cases is the durability risk you are carrying. In a market like Texas or Virginia in 2026, that gap is wide enough to flip a site ranking.

The siting consequence: rank candidate sites on pre-incentive fundamentals first, then let net-of-incentive economics break ties, never the reverse. A site that wins only because of a deep incentive sees its ranking invert the moment the incentive is clawed back or sunsets, and you cannot move a slab to chase a tax break that moved. Put the incentive in the assumptions register as a flagged, dated, risk-weighted figure (the same discipline applied to contested financial figures in Chapter 1.8), not as a reason the site exists. The 20–30 year asset must clear its hurdle on power, water, permitting and latency. The incentive is the margin of comfort on top, not the foundation.