AI Build Configurator Changelog

Pick any province and the model behaves exactly as it did in v1.2.0, in Canadian dollars. Pick any state and the whole model switches to US rules and US dollars at once. Nothing is converted at the end as a display trick. Each input is restated in USD before the maths runs, and the state rules are applied inside the calculation.

Currency

The model's capital benchmarks, accelerator prices and staffing constants were originally derived from USD-quoted sources at 1.38 Canadian dollars per US dollar. For a US state, every one of those constants is divided by the same 1.38, which recovers the underlying USD benchmark rather than layering a second exchange-rate assumption on top. For reference, the ECB-derived reference rate was 0.7216 US dollars per Canadian dollar on 11 September 2026, or about 1.386. Default revenue rates are converted the same way. If a user has already moved the revenue slider and then switches country, the setting is carried across at the same rate rather than reset.

Power rate and queue

Each state carries an indicative large-load industrial rate in USD per kilowatt-hour and an indicative queue in months. Rates start from each state's EIA industrial average, indexed to the April 2026 US industrial mean of 8.66 cents. In the 13 PJM states and DC, capacity cost is already inside that rate; PJM's 2028/2029 auction cleared at the cap of roughly $325 per MW-day. Transmission-voltage loads usually settle below the state average, so the slider stays live for a negotiated figure.

Sales and use tax on equipment

When a state has no data centre exemption open to new projects, the model applies that state's combined state and average local sales tax to the taxable share of each capital line. That share is 60% of fit-out, 75% of cooling, 85% of network fabric and 100% of accelerators. Shell and land are treated as real property and excluded. The tax appears as its own row and bar in the widget and its own line in the CSV. In 2026 this is no longer a theoretical line: Arizona's exemption is closed to new applications until 30 June 2029, Illinois paused its incentives for two years from 1 July 2026, and Ohio stopped approving new ones on 27 May 2026.

Large-load take-or-pay tariffs

Where a utility bills large loads on a minimum-demand basis, the model never lets billed IT energy fall below the floor multiplied by full-load IT energy once the site passes the tariff threshold. At 88% utilisation the floor rarely binds. At lower utilisation it does, and the shortfall is shown as its own line. A 50 MW neocloud in Virginia at 60% utilisation carries about $10.4 million a year of take-or-pay energy under the modelled GS-5 floor. Under pass-through billing, the line is marked as the tenant's cost.

Virginia uses the published 85% transmission and distribution floor, which is the more demanding of its two components. Floors of 80% are planning assumptions where a published blended figure is not directly comparable.

Construction cost, property tax and staffing

Shell, fit-out, cooling, network and staffing are multiplied by a state construction location factor, with the US national level at 1.00. The range runs from 0.86 in Arkansas and Mississippi to 1.30 in Hawaii. Property tax and insurance are scaled by a state index, from 0.60 in Delaware, Alabama and Hawaii to 1.40 in Illinois and New Jersey. Montana's special data centre property classification is reflected at 0.70.

The same builds, side by side

Default settings, 50 MW of IT load, 88% utilisation, B200-class accelerators. Canadian rows are in CAD and US rows in USD, so compare shapes across the border rather than raw figures.

The pattern is the point. For a neocloud that buys its own silicon, a closed sales tax exemption adds $125 to $160 million to a 50 MW build. That is a bigger swing than the gap between the cheapest and the median power rate. The full 51-row state table, with rate, queue, tax status, tariff and construction factor for every state, is on the AI data center economics page.

Equipment tax status as modelled for new projects

Exemption open (30 states): Alabama, Arkansas, Connecticut, Florida, Georgia, Idaho, Indiana, Iowa, Kansas, Kentucky, Louisiana, Maine, Maryland, Michigan, Minnesota, Mississippi, Missouri, Nevada, North Carolina, North Dakota, Pennsylvania, South Carolina, Tennessee, Texas, Utah, Virginia, Washington, West Virginia, Wisconsin and Wyoming.

Paused or closed to new projects, so tax is modelled (7): Arizona, Illinois, Massachusetts, Nebraska, New Jersey, Ohio and Oklahoma.

No data centre exemption, so tax is modelled (10): Alaska, California, Colorado, the District of Columbia, Hawaii, New Mexico, New York, Rhode Island, South Dakota and Vermont.

No state sales tax (4): Delaware, Montana, New Hampshire and Oregon.

1. Standard. All 51 US locations are live on your three lease-model configurations, in USD, with state tax and tariff rules applied. The clipboard summary now states the currency and, where relevant, the sales tax and large-load tariff. No action required.
2. Pro. Everything above across all seven configurations. The CSV export gains six rows. Country, Currency, Construction location factor, Large-load tariff and Take-or-pay energy cost follow the Utility row, and Sales and use tax on equipment follows Accelerator capex. Every currency column now reads CAD or USD to match the location. The rows also appear on Canadian exports with neutral values (CA, CAD, 1, n/a, 0).
3. Enterprise. Everything above. Pinned cost constants stay in CAD and are converted at 1.38 for US locations, then multiplied by the state location factor. If your pinned basis is already a US-dollar basis, tell us and we will set it per country. Custom US sites, utility territories or negotiated rates can be added per licence key.

State policy status reflects actions reported through August 2026. State power rates start from EIA state industrial averages indexed to the 2026 national industrial mean. They are planning constants, not month-specific tariff quotes. Where a tariff floor or threshold is a modelling assumption rather than a published figure, it is labelled as one above.

• U.S. Energy Information Administration, Electric Power Monthly, Table 5.6.A — state industrial average prices and the April 2026 national industrial average
• PJM Interconnection, 2028/2029 Base Residual Auction results, July 2026, and the capacity price cap and floor extended through the December 2026 auction
• Virginia State Corporation Commission, final order in Dominion Energy Virginia's 2025 biennial review — GS-5 rate class
• Public Utilities Commission of Ohio, AEP Ohio data centre tariff approved July 2025
• Texas Senate Bill 6 (2025) large-load interconnection and curtailment provisions, and the governor's 2026 direction to the PUCT
• State data centre incentive actions in 2026 as reported by the National Conference of State Legislatures, MultiState, Thomson Reuters Tax & Accounting and Telecompetitor — including the Arizona moratorium, Illinois, Ohio and Nebraska pauses, the Massachusetts closure, the New Jersey freeze, the Oklahoma sunset and the Washington changes of 1 July 2026
• Published combined state and average local sales tax rates
• Published construction city cost indices, rebased to US national = 1.00
• European Central Bank reference rates for CAD/USD, 11 September 2026

Nothing. That is not a throwaway line, it is the reason the product is structured the way it is.

The entire model — every cost constant, every provincial rate, every interconnection queue estimate, every accelerator specification — lives in one hosted JavaScript file on our server. A licensed embed is two elements on your page: a container div carrying your key, and a script tag pointing at our file. When we replace that file, your page serves the new model on its next load. There is no package to update, no build to run, no snippet to re-paste, and no version drift between licensees.

That matters more than it sounds. The failure mode for a cost calculator is not that it breaks. It is that it quietly keeps working while the numbers behind it go stale, and a consultant puts a five-year-old power rate in front of a client who knows better. Central hosting means that cannot happen to a licensee, because there is no stale copy to run.

The one thing worth checking is your own copy. If you have published prose alongside the widget that quotes specific outputs — a blog post citing a cost per megawatt, a proposal template with a payback figure, a case study with a capital number — that text is now out of step with the tool sitting next to it. The widget updated. Your paragraph did not. Search your site for hard-coded figures before your next client meeting.

Power is the recurring value of the subscription. It is also the part of the model that moves for reasons nobody controls. Three provinces changed in ways that alter site selection logic, not just arithmetic.

Quebec was repriced, and it is the biggest single change in the release

Hydro-Quebec filed a dedicated rate for large data centres with the Regie de l'energie averaging 13 cents per kilowatt hour — roughly double what large-power customers pay today. It applies to facilities above 5 megawatts, is expected in force in the second half of 2026 subject to approval, and applies automatically to all new customers in the segment. Sites already connected transition to it over five years.

Alongside it, the utility proposed moving the rate for cryptographic use applied to blockchains to an average of 19.5 cents per kilowatt hour, with a three-year transition for existing customers, on the reasoning that the activity is energy intensive with limited economic return to the province.

The model moved Quebec from 5.5 cents to 13.0 cents. The practical effect is blunt: run the hybrid mining and AI profile against Quebec now and it returns negative EBITDA. That is not a modelling artefact, it is the correct signal. Under a 19.5 cent blockchain tariff the hashing side of a hybrid site does not clear its own power cost. Quebec has spent two decades as the reflexive answer to "where should we put it," and for this class of load that answer has changed. Anyone still running a Quebec hybrid thesis on 2024 numbers is working from a document that a regulator has already overtaken.

British Columbia cancelled the rate schedule the old figure was based on

Rate Schedule 1823, the stepped transmission rate, closed to all customers at the end of the fiscal 2026 billing year, and any remaining customers were moved to Rate Schedule 1830. RS 1830 is now the default transmission service rate, and from 1 April 2026 its three separate energy charges were consolidated into a single charge. The current structure is 4.914 cents per kilowatt hour on all energy, plus a demand charge of $12.178 per kVA of billing demand. Bills across the board rose roughly 3.75 percent on the same date.

For a high load factor data centre, that demand charge spreads across a lot of kilowatt hours and blends out near 6.6 cents. That is the figure the model now carries, up from 6.2 cents, and it is derived from the published tariff components rather than a rounded industry estimate.

The queue estimate moved more than the rate did. BC Hydro's process for emerging industries is a competitive allocation, not a queue — applicants compete for a finite block of capacity rather than waiting their turn. The utility received 15 applications representing close to 800 megawatts of requested load, with successful applicants notified from mid-September 2026. The model now carries 36 months for BC and flags the region as allocation-constrained, because the honest answer to "how long is the wait" in British Columbia is that there may not be a wait, there may be a refusal.

Alberta is no longer the fast lane, and v1.0 said it was

This is the province-wide change that most affects how the tool advises. Alberta spent years as the fastest interconnection path in Canada, and the model carried a 14-month queue on that basis. Provincial policy has since moved decisively.

The Alberta Electric System Operator's Phase 1 large-load allocation was capped at approximately 1,200 megawatts, and that cap was fully allocated to two projects with in-service dates in 2027 and 2028. Meanwhile the connection project list still carries roughly 19,500 megawatts of requested data load across dozens of projects. Phase 2 of the Large Load Integration programme is developing the long-term framework covering connections, planning, operations, markets and tariff, without a published completion timeline. In June 2026 the province filed a Data Centre Regulation creating a framework for loads of 75 megawatts and above, and the system operator has been building out a bring-your-own-generation pathway as the practical route for capacity beyond the interim cap.

So the queue went from 14 months to 26, and the region flag changed from fast to byog. Alberta is still a serious answer — but the answer now involves generation you bring yourself, regulatory approval for that generation, and emissions permitting, not simply a faster interconnect.

Every province, before and after

Rates are indicative blended all-in industrial figures in Canadian dollars per kilowatt hour at high load factor, for screening purposes. They are not tariff quotes. Every licensee should verify against the applicable rate schedule and their own load profile before a rate reaches a client document.

Construction costs moved in one direction across every published index we track, and the model moved with them. The global average shell-and-core construction cost for a single-tenant facility reached roughly $11.3 million US per megawatt in 2026, up from $7.7 million in 2020 — a 47 percent increase in six years that has almost nothing to do with general inflation and almost everything to do with electrical density, liquid cooling, redundant power trains and testing regimes. AI-optimised facilities run materially higher, with all-in figures for a deliverable megawatt commonly quoted between $20 million and $37 million US.

Every profile's shell, fit-out, land, utility contribution and network lines were raised. Soft costs moved from 8 percent of hard cost to 9 percent. The stabilised-asset valuation multiple came down from 15 times EBITDA to 14, reflecting deal multiples compressing through 2026 even as transaction volume set records.

Why every build timeline got longer

Not because construction got slower. Because procurement did. Medium-voltage switchgear has been running lead times around 22 months. High-capacity transformers sit near 18. The average across critical components is longer still. Against that backdrop, a meaningful majority of data centre projects in 2025 slipped at least three months from their original schedule.

The build duration figures in the model already assume competent procurement rather than luck, and they are added to the provincial queue to produce a single time-to-first-revenue number. Version 1.2.0 also exposes the underlying lead time assumptions through the public API so that enterprise licensees integrating the engine into their own reporting can sanity-check a schedule a client brings them against the same figures the model uses.

This change tracks a step-change in rack power across the industry rather than a refinement of an estimate.

Version 1.0 capped direct-to-chip liquid cooling at 140 kilowatts per rack, reflecting a GB200-class reference deployment. The industry has moved well past that ceiling. GB300 NVL72 sits at roughly 140 kilowatts per rack on its own, and the Vera Rubin VR200 NVL72 ships in two power profiles — Max Q at roughly 1.8 kilowatts per GPU and 190 kilowatts per rack, and Max P at roughly 2.3 kilowatts per GPU and 230 kilowatts per rack. Vendor rack specifications for Vera Rubin pods quote rack power consumption up to 240 kilowatts. A 140 kilowatt ceiling cannot represent the hardware now being ordered.

Density ceilings were raised accordingly, and the incremental cooling capital cost was raised with them. Liquid cooling typically adds somewhere in the range of 7 to 10 percent to total build cost versus a comparable air-cooled facility, and it re-weights the budget — mechanical systems rise from roughly 22 percent of facility cost to roughly 33 percent while electrical falls. Expressed per megawatt, the liquid premium lands in the low single-digit millions.

The air-cooled PUE moved upward deliberately, tracking measured industry performance. Facilities at 20 megawatts and above averaged a PUE around 1.44 globally in 2025, so the model now sits closer to that observed figure and reports a power bill reflecting how these halls actually run.

The accelerator table, rebuilt

Two existing keys were repointed at current silicon rather than removed, so no embed anywhere breaks. Two new classes were added. Unit costs are now expressed as fully loaded cost per accelerator — the GPU plus its share of CPU, NIC, storage, chassis and rack-scale interconnect — derived from published system and rack prices, which is the number that actually hits a capital plan.

Rubin-generation pricing is the softest data in the release and is labelled as such in the source file. Channel quotes for a VR200 NVL72 rack have circulated between roughly $5 million and $7 million US, while a widely cited May 2026 sell-side estimate put the figure nearer $7.8 million, with memory alone accounting for around a quarter of system cost after a very large increase in high-bandwidth memory pricing. The model carries a figure inside that band. Licensees modelling a Rubin deployment against a real quote should override it.

Operating expense reporting across the sector has become considerably more granular through 2026. Lease disclosures, operator filings and colocation cost surveys now routinely separate facility operating cost from the carrying cost of the asset and the overhead of the business running it. Version 1.2.0 brings the model's operating expense breakdown into line with that convention.

Version 1.0 grouped operating expense into three lines: electricity, maintenance as a percentage of capital, and staffing. Version 1.2.0 adds a fourth covering property tax, insurance and general administration, set per profile as a per-megawatt charge plus a percentage of revenue. It appears in the widget's operating expense breakdown, in the clipboard summary, and as two separate rows in the CSV export.

The effect on reported margin varies by commercial structure, and that variation is the useful part. A powered shell with power passed through to the tenant genuinely carries very low operating cost against its rent — one publicly disclosed figure from a large lease put pure facility operating cost near $15 US per kilowatt per month against wholesale rent in the high $100s. Structures that own silicon, employ technical staff and sell compute hours carry materially more. Breaking the fourth line out lets the model show that spread across the seven profiles rather than flattening it into a single number.

Payback periods lengthened correspondingly. A 50 megawatt powered shell in British Columbia moved from roughly 8.7 years to roughly 10.2. A 20 megawatt neocloud moved from 3.6 years to 4.2.

Licensees who have quoted a margin or payback figure from the previous model inside a live document should refresh it against the current output. The revised range reflects both the expanded expense breakdown and the capital benchmark movement described above, and it is the range that will stand up in a room with a chief financial officer in it.

1. Standard. All revised provincial data, all revised capital benchmarks, the new operating expense line and the new accelerator classes are live in your three lease-model configurations. Your profile list is unchanged. Clipboard summary now includes the new operating expense figures. No action required.
2. Pro. Everything above, across all seven configurations, plus two additional columns in the CSV export breaking out property tax and insurance separately from general and administrative cost. If you have built a downstream template or spreadsheet that parses our CSV by row position rather than by label, check it — two rows were inserted into the operating expense block.
3. Enterprise. Everything above, plus the newly exposed lead-time constants through the public API for firms integrating the engine into their own reporting. If your cost constants are pinned to an internally agreed basis, your pinned values are unaffected by the benchmark revisions — but the new operating expense category will apply unless you override it, and the density ceilings and accelerator classes have moved underneath your pins. Worth a review call.

Provincial rate data is taken from published utility tariffs and regulatory filings wherever a published figure exists, and from the utility's own rate design documentation where it does not. Construction benchmarks come from published 2026 industry cost indices. Accelerator specifications come from vendor and system-integrator published rack specifications. Pricing for shipping silicon comes from published rate cards and system prices; pricing for pre-volume silicon is a band, and is labelled as a band in the source file.

• BC Hydro transmission rate schedules, fiscal 2027 updates effective 1 April 2026 — RS 1823 conclusion, RS 1830 energy and demand charges
• BC Hydro emerging industries connections process — application volume and notification timing
• Hydro-Quebec filing to the Regie de l'energie, February 2026 — proposed large data centre rate and cryptographic use rate
• Alberta Electric System Operator Large Load Integration programme materials and connection project reporting
• Alberta Data Centre Regulation, filed 9 June 2026
• Published 2026 global data centre construction cost indices for shell-and-core and AI-optimised per-megawatt benchmarks
• Published cost-structure analysis on the air-cooled to liquid-cooled budget shift
• Published operator and vendor rack specifications for GB300 NVL72 and Vera Rubin VR200 NVL72 power profiles
• Published cloud GPU rate cards and system pricing for accelerator cost and revenue calibration
• Published wholesale colocation asking-rate surveys for North American primary markets

Every figure in the model is a planning-grade screening estimate. Rates, queues, allocation policy and silicon pricing change without notice. Nothing on this page or in the widget is an engineering estimate, a utility commitment, a price quote, or financial advice.