Anthropic keeps building larger AI data centers, and its power bill grows with them.
An article published by Better Data Center Project on November 16, 2026, titled How AI is powered: A breakdown of BYOP and grid power, provides a forecast: by 2028, 87% of the electricity Anthropic uses is expected to come from natural gas and 11% from coal.
This is neither an Anthropic-disclosed energy commitment nor a record of actual consumption. It is a forecasting model presented in the article.
Still, the model highlights an overlooked fact: Claude’s expansion roadmap is moving into the territory of conventional energy and grid infrastructure.
For developers, the practical questions are not limited to how clean data centers really are. They are:
- Can the Claude API expand capacity as planned?
- Can inference prices continue to fall?
- Will capacity be throttled during peak demand if power is insufficient?
What 87% Natural Gas Means
First, the boundaries of the data need to be clear.
The article projects that by 2028, 87% of the electricity Anthropic uses will come from natural gas and 11% from coal. “Electricity used” refers to the power supply supporting its AI infrastructure, not the everyday electricity consumption of Anthropic’s offices.
Natural gas becoming a practical choice is not difficult to understand.
AI data centers operate under sustained, heavy loads. Training jobs can wait in a queue, but inference services are generally expected to remain online around the clock. They need more than a quantity of electricity; they need stable voltage, adequate capacity, and predictable delivery.
Natural gas generation offers several practical advantages:
- Generating units can be dispatched relatively easily, making them suitable for variable loads;
- They can provide peak balancing when combined with renewable energy;
- They generally emit less than standalone coal plants;
- For some power developers, natural gas remains a mature source that can support project financing and long-term contracts.
But 87% is not a casual figure. Under the article’s forecast, most of Anthropic’s power supply would be tied to natural gas generation. Gas prices, pipeline access, gas-turbine delivery schedules, and local air-quality rules would all become variables in Claude’s infrastructure.
The 11% share for coal cannot be dismissed either.
Coal generation usually has a higher carbon intensity than natural gas, and its equipment often has long operating lives. If that share comes from existing generating units, Anthropic would be purchasing a mixed grid supply rather than dedicating a gas turbine to every GPU.
Two concepts need to be separated:
- Power mix describes where the electricity comes from.
- Grid interconnection describes how a data center connects to those sources.
The two are related but are not the same metric. An 87% natural gas share does not mean that 87% of Anthropic’s data centers would be powered by company-owned gas plants. Nor does the fact that the electricity comes through the grid eliminate coal’s effect on the capacity and stability of local grids.
BYOP Is More Than Bringing Your Own Power
BYOP stands for Build Your Own Power. More precisely, it refers to enterprises that build and operate their own data centers and supporting power infrastructure rather than relying entirely on the public grid.
The article distinguishes two models for AI compute infrastructure:
- BYOP: building and operating the data center;
- Grid-powered infrastructure: leasing or building compute facilities whose electricity is supplied by the local utility system.
“Building your own data center” can sound like buying a few rows of racks. In practice, procuring power, selecting a site, and securing grid interconnection are three separate challenges.
Procuring Power
A company may sign a long-term power-purchase agreement, or PPA, with a utility, an independent power producer, or an energy trader. Such an agreement locks in prices, delivery volumes, and contract terms; it does not move a generator into the server room.
Long-term PPAs can reduce exposure to price volatility, but they introduce different risks. A project may not yet be connected to the grid, equipment may arrive late, and minimum purchase commitments may not align with actual training demand.
Site Selection
Choosing a site involves more than land and electricity prices.
Companies must also assess:
- whether connectable generating capacity is available locally;
- whether high-voltage lines and substations can accommodate additional load;
- whether cooling water is sufficient;
- whether fiber networks can connect the facility to other training clusters;
- whether natural gas pipelines, coal plants, or renewable projects have clear delivery schedules.
Large cloud providers such as Google, Microsoft, and Amazon can draw on data center footprints, procurement channels, and financial scale accumulated over many years. If a startup copies that playbook, transmission capacity is likely to be its first constraint—not model code.
Grid Interconnection
This is the step most likely to cause delays.
A 100 MW data center is not an ordinary commercial project. Interconnection requires the grid operator to assess lines, transformers, protection systems, and regional loads. The construction schedule proposed by the engineering team may also exceed the standard delivery period for electrical equipment.
BYOP therefore does not mean simply “bringing your own power.” It is closer to a capital-intensive portfolio in which a company must manage IT, mechanical and electrical systems, fuel supply, regulation, and grid relationships at the same time.
The Cost Model Must Be Recalculated
A common formula for discussing LLM costs is:
Total cost = GPU purchase and depreciation + data center + power + networking + operations
The formula is correct, but too coarse.
Breaking a model service down further gives:
Cost per unit token = (GPU depreciation + power + data center + networking + operations) / effective tokens
The real difference comes from “effective tokens.”
Suppose a rack has already been purchased, but customer requests occupy only part of its capacity for the entire day. Unused GPUs still incur depreciation, cooling, and maintenance costs. Higher utilization spreads fixed costs over more output; lower utilization makes seemingly inexpensive hardware expensive.
Data center electricity costs are also not simply kilowatt-hours multiplied by a unit price. They include:
- grid capacity charges and transmission and distribution fees;
- backup generators, fuel, and scheduled test runs;
- energy losses associated with UPS systems;
- electricity consumed by cooling systems;
- premium costs during extreme weather or grid congestion.
A long-term PPA can lock in some price and delivery risk in advance. But locking in a price does not lock in capacity. If a project enters service late, the model service will still need to buy expensive grid power in the meantime.
At another level, model architecture can also change the energy bill. Mixture-of-Experts models typically activate only a subset of parameters; quantization can reduce memory and computation requirements; caching can reuse repeated prefixes; and continuous batching can improve GPU utilization.
Ultimately, these technical and commercial decisions reach the infrastructure layer: less idle compute, lower energy consumption per token, or greater concurrency density.
That is why Claude’s long-term unit cost cannot be assessed from its API list price alone.
List price is a revenue-side variable. Power, site selection, grid interconnection, and backup capacity determine whether the service can continue receiving enough supply and whether the company can keep reducing prices.
The Controversy Goes Beyond Carbon Emissions
When Anthropic’s forecast data reached the front page of Hacker News, it had earned 176 points and 234 comments.
The discussion quickly moved beyond whether AI is environmentally friendly and toward more specific questions: Is there enough public data to support the 87% and 11% figures? Why was coal not excluded entirely? What happens when gas prices rise? Could local grid congestion distort the expansion forecast?
Natural gas generation is not inherently low-carbon.
Its emissions depend on upstream methane leaks, plant efficiency, fuel sourcing, and operating patterns. If the local grid relies primarily on coal generation, marginal power and annual average power are also different concepts.
For Anthropic, however, coal may have one very practical benefit: providing a degree of hedge.
Natural gas is constrained by pipelines and fuel prices, while coal generation is affected by coal supplies, plant maintenance, and cooling conditions. A mix of fuels and locations may reduce the risk of disruption tied to any single energy source.
The truly difficult issue is the local grid.
A region may appear to have sufficient annual generation without enough load capacity at the substation serving the planned data center. Electricity produced hundreds of kilometers away may not reach the racks when it is needed.
Developers therefore care less about abstract energy slogans and more about whether Claude can expand reliably:
- Will new regions be delayed by the grid interconnection process?
- Will additional power be available when traffic surges?
- Will inference pricing come under pressure after natural gas or coal prices rise?
- How will services be prioritized during grid curtailments or fuel supply disruptions?
These questions map directly to API reliability, latency, and pricing.
Why the Developer Community Is Debating It
The most prominent criticism on Hacker News centered on whether the forecasting methodology was transparent enough.
The key facts currently available are that the article was published on November 16, 2026, its original title was How AI is powered: A breakdown of BYOP and grid power, and it distinguished between BYOP and grid-powered infrastructure.
But 87% natural gas and 11% coal are projections, not published operational data. Readers need to ask which corporate disclosures, procurement agreements, and regional grid data fed the model. Are there undisclosed energy contracts? Was the mix calculated using generation, contracted capacity, or actual electricity supplied?
These details could change the conclusions.
Another line of criticism concerned the difference between direct power sourcing and grid purchases. Under BYOP, a company may have clearer renewable-energy contracts and greater control over site selection. Grid-powered infrastructure spreads construction schedules and fuel risks across utilities, but it also adds uncertainty from regional congestion and queue-based interconnection.
These two models should not be reduced to a ranking of which is more advanced.
For hyperscale operators with stable long-term demand and the capital to manage it, BYOP can place power generation, data centers, and model training on a shared schedule. For teams whose demand is still being validated and whose scale may change rapidly, locking in heavy assets early can be dangerous.
The most useful part of the debate is separating three layers:
- Public facts: the article’s publication date and title, and the two infrastructure models it distinguishes;
- Model projections: a predicted 2028 energy mix of 87% natural gas and 11% coal;
- Author analysis: how the power mix maps to cost, emissions, and expansion risk.
Mixing these layers makes every energy conclusion appear more certain than it really is.
Original article: Better Data Center Project, How AI is powered: A breakdown of BYOP and grid power; Hacker News discussion: Item 50597612.
Editorial Takeaways
These numbers matter not because they prove that Anthropic’s AI expansion is environmentally harmful, but because they return model competition to the power system. GPUs can be upgraded every quarter; transmission lines, substations, and gas turbines may take years. The larger Claude becomes, the harder it is for software iteration to conceal the infrastructure timeline. AI Agents designed to run for long periods will inherit the same capacity constraints.
The more consequential shift may be BYOP’s departure from the traditional model of purchasing cloud capacity. Companies may participate directly in power generation, data center construction, and grid-interconnection arrangements rather than simply buying tokens from a cloud provider. Model companies are beginning to manage long-term contracts much like energy-intensive manufacturers. That can improve supply certainty, but it also leaves fuel-price and regulatory risks on their balance sheets.
Even so, 87% and 11% remain projections rather than disclosed facts. Whether the underlying model includes undisclosed PPAs and how it defines regional grids could alter the percentages. At this stage, one conclusion is clear: “reliable power” must be added to the list of Claude’s potential bottlenecks.