AI infrastructure spending is accelerating across model usage, data centers, storage, and connectivity, but enterprise cost controls are not keeping pace. For CTOs, platform leaders, FinOps owners, and finance teams, the issue is increasingly one of measurement: costs can arrive after work is done, and related infrastructure charges may sit outside a single AI invoice.
What changed
The news hook is a physical buildout problem as much as a software one. Associated Press reporting says a rapid nationwide expansion of AI-driven data centers is triggering local resistance in Florida, where more than 20 counties and municipalities have voted to reject major data centers and others have banned construction for at least a year.
That backlash is already producing policy responses. The AP reported that Florida Gov. Ron DeSantis signed legislation in May requiring large-scale data centers to pay their own energy and water costs, and that Byron Donalds announced federal legislation requiring such centers to source electricity and water privately rather than push costs onto utility customers.
At the enterprise budget level, spending pressure is rising just as fast. AI Business, in an article labeled as sponsored by Google Cloud, citing Gartner, said worldwide AI spending could reach $2.6 trillion this year, implying 47% year-over-year growth.
Why B2B teams should care
The core problem is not that AI is expensive, but that costs are known too late and described too loosely for planning. AI Business’ sponsored analysis said opaque pricing and backward-looking bills are pushing enterprises to rethink AI strategy, since users often learn costs only after a model finishes the work.
That creates a mismatch between experimentation speed and financial control. Starburst CEO Justin Borgman said his company runs most AI production work on Anthropic, that Claude and other tools write about a third of Starburst’s code, and that Starburst does not have good predictive modeling of spend today. The same article says the company’s idea-to-production time has dropped by up to 60%, which helps explain why usage can scale before forecasting matures.
Finance teams face the same issue at the billing layer: Pega COO and CFO Ken Stillwell described reasoning-token costs as a “black box,” a shorthand for why usage charges are hard to forecast before invoices arrive. See our OpenAI Pricing Guide: ChatGPT Plans and API Usage Controls for a similar forecasting problem in usage-based AI billing.
Who is affected
Several enterprise functions sit inside this gap. Platform and infrastructure teams often manage GPU or compute capacity; FinOps then has to attribute the resulting spend. Procurement may be negotiating vendor-managed compute or reserved capacity, while finance leaders remain exposed to storage, networking, and utility costs that often fall outside direct model usage charges.
The infrastructure categories extend beyond tokens. Citing Seagate materials, a market note said enterprises and cloud providers deploying more GPUs also require significant storage capacity, linking high-capacity nearline drives to AI training and large language models, and noting that Seagate often cites multi-terabyte and multi-petabyte configurations in its marketing and technical documentation.
Connectivity is another cost layer. A market commentary from Kalkine cited company disclosures showing Megaport announced AI contracts in May and June 2026 worth a combined A$700m+, with 983 enabled data centres at the close of FY25 and mid-2026 materials putting that footprint above 1,100 across 31 countries.
What teams should check now
Teams evaluating AI infrastructure cost management should break total cost into separate lines, not treat model bills as the whole picture.
- Model and token charges: Review whether pricing is transparent enough to forecast before workloads run, especially where reasoning-token charges are hard to estimate in advance, as Pega’s Ken Stillwell described.
- GPU or compute capacity: Check where capacity commitments are being made by platform teams, even when spending visibility is still backward-looking.
- Storage: Account for storage growth alongside GPU rollout, per the Seagate-sourced figures above.
- Networking and connectivity: Include private interconnection and network expansion costs tied to AI workloads, per the Megaport figures above.
- Energy and water exposure: Review whether existing contracts already pass utility costs through to data-center operators, or whether that exposure still sits with the enterprise.
- Regulatory tracking: Track Florida’s May law requiring data centers to pay their own energy and water costs, and Byron Donalds’ proposed federal legislation on utility-cost pass-throughs.
Teams should also check whether they rely on backward-looking bills rather than predictive forecasting, the gap AI Business and Starburst’s Borgman both described above.
What remains unclear
- Not yet confirmed: a clean enterprise benchmark for average AI compute overspend, or how many enterprises cannot accurately attribute costs across tokens, storage, networking, and utilities.
- Not yet confirmed: the full methodology behind AI Business’ cited Gartner forecast of $2.6 trillion in worldwide AI spending this year.
- Not yet confirmed: whether Donalds’ proposed federal legislation will advance; the AP report includes no bill number or legislative text.
- Not yet confirmed: how representative the Megaport and Seagate figures are for broader enterprise economics, given AI Business’ Google Cloud sponsorship and the lower-tier sourcing behind those items.
What to watch next
The first near-term signal is whether utility-cost regulation spreads beyond Florida after DeSantis’ May action and Donalds’ federal proposal. The second is whether local resistance intensifies; the AP report already points to more than 20 Florida counties and municipalities rejecting major facilities and other jurisdictions imposing year-long construction bans.
The third signal is enterprise disclosure: buyers should watch for companies to publish more granular controls around forecasting, attribution, and non-model cost categories, especially near compute capacity lock-in like the one discussed in Reflection Nebius Compute Deal Signals AI Capacity Lock-In. The visibility gap may become a larger operational issue than model pricing itself if infrastructure costs keep spreading across vendors and billing systems.
Sources
- Audacy / The Associated Press, Byron Donalds proposes legislation to stop AI data centers from driving up utility costs
- AI Business, As AI spending climbs, enterprises are serious about token cost
- Kalkine, Megaport (ASX:MP1) caught in the tech sell-off as investors question the AI infrastructure boom
- AD HOC NEWS, Seagate stock trades steady as data storage demand and AI infrastructure spending support outlook
