Two futures and a plug

The first question for an AI data center is no longer which GPU to buy.

It's how many megawatts you have. And when they get connected.

I see two possible futures for AI this decade. They are opposite in almost everything except one: both run into the same limit. It isn't silicon. It's the plug.

I went to Epoch AI for the numbers. It's the research group that measures this best. Its data is public under a CC BY 4.0 license, and I redrew its series for this piece. Every figure carries its data table and sources underneath. Where the conclusion is mine, I say so.

The pace

In 2020, training GPT-3 took about 5 MW of power. In 2023, GPT-4 took about 20. In 2025, Grok 3, about 110. And in September 2026, GPT-6 Astra took at least 233 MW: a lower bound, because it was trained on at least 100,000 GB200 chips. Epoch estimates that the power of the largest training runs grows 2.1x every year.

Figure 1

The largest training run already draws at least 233 MW; Epoch projects 4 to 16 GW by 2030

Estimated cluster power of record-setting training runs since 2020, among those Epoch estimates (MW, log scale), and the 2030 projection.

Show the data

Series	Date	Value	Note
Record training power draw	May 2020	5.1 MW	GPT-3
Record training power draw	Mar 2023	20 MW	GPT-4
Record training power draw	Dec 2023	38 MW	Gemini 1.0 Ultra. Speculative estimate per Epoch.
Record training power draw	Feb 2025	110 MW	Grok 3
Record training power draw	Sep 2026	233 MW	GPT-6 Astra (lower bound). Trained on at least 100,000 GB200s: the figure understates the real draw.
Record training power draw	Jul 2030	4 GW to 16 GW	2030: 4 to 16 GW. Band: Epoch and EPRI (Aug 2025), 4 to 16 GW per run. Point: 'AI in 2030' (Sep 2025), about 10 GW if the trend continues.

If the trend holds, Epoch and EPRI, the US power industry's research institute, project that the largest training run of 2030 will need between 4 and 16 GW. Epoch's 2030 report puts it at about 10 GW if the trend continues: more than all but the largest power plants in the world produce. For a single run.

And it isn't just training. The power capacity of the world's AI data centers went from 0.9 GW at the end of 2022 to 13.5 GW at the end of 2024 and about 30 GW at the end of 2025, according to Epoch's extrapolation from chip sales. Comparable to New York State's peak demand.

Why it isn't the hardware

Intuition says chips keep getting better and that fixes everything. The numbers say otherwise.

Figure 2

Compute demand grows 5x a year; GPU energy efficiency, 1.34x

Annual growth multiplier per Epoch AI. Demand in blue, what hardware delivers in ochre.

• Demand
• Hardware

The dashed line marks 1x: no growth.

Show the data

Item	Group	Value	Note
Frontier training compute	Demand	5×	Frontier language models, since 2020. 90% confidence interval: 4 to 6.
Power of the largest training runs	Demand	2.1×	Top 10 by compute, since 2010. 90% interval: 1.9 to 2.2. Epoch and EPRI (2025) put it at 2.2.
Chip performance per dollar	Hardware	1.49×	Since Q1 2023, in constant dollars. 90% interval: 1.36 to 1.66.
GPU energy efficiency	Hardware	1.34×	FLOP/s per watt, since 2008: doubles every 2.4 years. 90% interval: 1.29 to 1.39.

The compute used to train frontier language models has grown about 5x a year since 2020. GPU energy efficiency, measured in operations per second per watt, improves 1.34x a year since 2008 (1.4x for the most advanced AI hardware). Chip performance per dollar, 1.49x. Epoch puts it this way: AI hardware efficiency improves around 40% a year, and compute demand eats it.

There is one detail I think sits at the center of all this. According to an analysis by four Epoch researchers, chips cost about 10 times more than the electricity they use. And in the frontier models Epoch analysed in 2024, energy was only 2% to 6% of the cost of developing them.

In other words: energy isn't expensive. It's slow.

You buy chips with money. You buy connected power with time.

In the US, the median power project that came online in 2025 took 61 months from requesting a connection to operating. In 2015 it was 36. Some 2,061 GW are waiting in interconnection queues, and only 13% of what requested a connection between 2000 and 2020 ever got built. GE Vernova holds 122 GW of gas turbine orders and slot reservations.

When something is scarce, it gets a price. In PJM, the largest power grid in the US, it already has one:

Figure 3

In PJM, the capacity price went from $28.92 to between $325 and $333 and has sat at the cap for three auctions

Clearing price of the capacity auction of PJM, the largest US grid (dollars per MW-day), by delivery year.

Show the data

Delivery year	Value	Note
21/22	$140
22/23	$50
23/24	$34.13
24/25	$28.92	The lowest price in the series shown.
25/26	$269.92	9.3 times the previous auction.
26/27	$329.17	At the regulated cap.
27/28	$333.44	At the cap. PJM attributes nearly 5,100 of the 5,250 MW of forecast peak growth to data centers (December 2025 release); per its July 2026 report, it cleared 6,516 MW short of the reliability requirement.
28/29	$325	At the cap; 6,831 MW short of the reliability requirement.

From $28.92 per MW-day to $269.92 in a single auction, and up to the regulated cap in the three after that. PJM attributes nearly 5,100 of the 5,250 MW forecast increase in peak demand for 2027/28 to data centers. And the 2028/29 auction cleared at the cap again, 6,831 MW short of what PJM needs to guarantee supply.

In fairness, Epoch doesn't see this as a wall. In the summary of its 2030 report it writes that electrical power "is unlikely to be a bottleneck before 2028, and seems solvable even after that". Those four researchers argue that the power problem will add months, not years, to build-out times, and that even their best projection for the US in 2030, an aggressive scenario of about 100 GW, can be met with gas, solar plus batteries and demand flexibility. And Epoch documents how it is being routed around: with gas generated on site, as at least 3 of the 7 Stargate sites in the US will do.

They have the data to say so. I read the same data differently. "Solvable" means building generation and grid for demand growth the US hasn't seen since the 1980s, by those same researchers' account. With projects that now take a median of 61 months from requesting a connection to operating. Meanwhile, someone pays the difference. In PJM, they already are.

And it isn't only the US. In Spain, the grid operator Red Eléctrica has granted 6,329 MW of transmission access to data centers, 34.2% of all granted demand. None of them is in service yet.

The paradox: intelligence to produce energy

This is the part I find most interesting. To produce more energy we need more intelligence. And to produce more intelligence we need more energy.

AI is already working on the first side of that loop. Google DeepMind is helping Commonwealth Fusion Systems optimize the SPARC fusion reactor, which CFS said was about 75% built in April 2026. In April 2026, PJM was using an AI tool to review the 811 applications in its new connection cycle, totalling 220 GW. And in a US Department of Energy demo, an AI drafted in one day a 208-page chapter of a nuclear license application that takes a team four to six weeks.

None of that has produced a new megawatt yet. Fusion has not delivered net electricity: Helion announced 150 million degrees and measurable fusion in February 2026 and keeps 2028 as its target for starting its plant for Microsoft. And two Epoch economists warn that automating R&D alone adds far less than automating the economy as a whole: R&D explains only about 20% of US labor productivity growth since 1988.

So the loop exists, but it turns slowly. And loops have a direction: whoever has the energy first, turns first.

With that in mind, the two futures.

Future one: two companies

In this future, two companies are left: Anthropic and OpenAI. Everyone else rents intelligence from them, with subscriptions at least twenty times pricier than today's.

It sounds like an exaggeration, but the concentration is already here. According to the data Epoch compiles, Anthropic's annualized revenue went from about $1 billion at the end of 2024 to $9 billion at the end of 2025 and, according to Bloomberg, to $65 billion by the end of July 2026. OpenAI's went from $5.5 billion at the end of 2024 to about $50 billion by the end of September 2026. The two companies count revenue differently, so the figures don't compare one to one. And according to a Menlo Ventures survey (Menlo is an Anthropic investor) of about 400 large US companies, the two already account for 86% of their spending on language model APIs.

At the top end, what people pay no longer looks like a $20 subscription:

Figure 4

ChatGPT's most expensive plan went from $20 to $500 a month; Claude's, from $20 to $200. The $20 plan stayed put

Monthly price of the most expensive individual plan, stepped at the date of each change (dollars).

• $20 plan (ChatGPT Plus and Claude Pro)
• Claude
• ChatGPT

Show the data

Series	Date	Value	Note
$20 plan (ChatGPT Plus and Claude Pro)	Feb 2023	$20	ChatGPT Plus and, since September 2023, Claude Pro.
Claude	Sep 2023	$20	Claude Pro.
Claude	Apr 2025	$200	Max 20x
Claude	Oct 2026	$200	Max 20x is renamed Ultra, same price.
ChatGPT	Feb 2023	$20	ChatGPT Plus.
ChatGPT	Dec 2024	$200	Pro
ChatGPT	Sep 2026	$500	Pro 500

ChatGPT's most expensive consumer plan went from $20 a month in 2023 to $200 at the end of 2024 and $500 in September 2026. Claude's, from $20 to $200. The $20 plan hasn't moved, and the cost of a fixed level of AI capability falls about 13x a year. What was frontier yesterday gets cheap fast. But the frontier moves, and staying on it costs more and more: ChatGPT's top plan is now 25 times the $20 Plus.

Where the "20x" already shows is in agents. Epoch estimates that an agent costs between $15.50 and $50 per hour of work at API prices, with $30 as the reference. I ran the numbers as an illustration: a full-time agent, 176 hours a month, would cost about $5,280 a month. Twenty-six times a $200 plan. Two hundred and sixty-four times a $20 one. It isn't the same product as a chat. It's what it costs to put AI to work.

And energy? In the risk factors of its IPO filing, Anthropic writes that "the availability of sufficient electrical power has emerged as a significant constraint on data center development and expansion in many geographies, including the United States" (S-1 filing).

An Epoch analyst estimates that at the end of 2025, OpenAI and Anthropic used only 15% to 20% of the world's AI compute. If they keep growing faster than everyone else, in a naive extrapolation he himself presents as an illustration, they would be near 80% in five years. And from there, he writes, further scaling "would require an economic transformation".

That's the key to future one. If energy is the limit, whoever has it under contract collects a scarcity rent. Two Epoch authors think inference demand is growing faster than capacity, even though capacity more than triples every year, and that a crunch is coming that would raise token prices for large models. Their crunch is about compute, not energy: the near-term brake they name is memory bandwidth. If they're right, and energy tightens on top of that, the price of the frontier doesn't fall. It rises.

Future two: the data center as a store of value

The opposite future: every company has its own AI, in its own data center, and that data center becomes the most valuable thing on its balance sheet.

This one has data behind it too. Five big tech companies (Google, Microsoft, Amazon, Meta and Oracle) owned about 70% of the world's AI compute at the end of 2025. And a growing share of the cash they generate goes into capex: about three quarters over the last four quarters in the series.

Figure 5

Hyperscaler capex went from 56% to 94% of the cash they generate (Q1 2022 vs Q1 2026)

Billions of dollars per quarter: cash capex versus operating cash flow of Microsoft, Amazon, Alphabet, Meta and Oracle.

• Operating cash flow
• Cash capex

Show the data

Series	Date	Value	Note
Operating cash flow	Q1 2022	$65.6B
Operating cash flow	Q2 2022	$69.2B
Operating cash flow	Q3 2022	$74B
Operating cash flow	Q4 2022	$79.3B
Operating cash flow	Q1 2023	$71B
Operating cash flow	Q2 2023	$96.9B
Operating cash flow	Q3 2023	$109.8B
Operating cash flow	Q4 2023	$99.8B
Operating cash flow	Q1 2024	$104.5B
Operating cash flow	Q2 2024	$114.6B
Operating cash flow	Q3 2024	$123B
Operating cash flow	Q4 2024	$136.3B
Operating cash flow	Q1 2025	$120.2B
Operating cash flow	Q2 2025	$134.6B
Operating cash flow	Q3 2025	$167.1B
Operating cash flow	Q4 2025	$180.9B
Operating cash flow	Q1 2026	$157.9B
Cash capex	Q1 2022	$36.6B	56% of operating cash flow.
Cash capex	Q2 2022	$38.4B	55% of operating cash flow.
Cash capex	Q3 2022	$41B	55% of operating cash flow.
Cash capex	Q4 2022	$41.9B	53% of operating cash flow.
Cash capex	Q1 2023	$36.6B	52% of operating cash flow.
Cash capex	Q2 2023	$35.4B	37% of operating cash flow.
Cash capex	Q3 2023	$38.3B	35% of operating cash flow.
Cash capex	Q4 2023	$44.1B	44% of operating cash flow.
Cash capex	Q1 2024	$46B	44% of operating cash flow.
Cash capex	Q2 2024	$55.6B	49% of operating cash flow.
Cash capex	Q3 2024	$61.2B	50% of operating cash flow.
Cash capex	Q4 2024	$76.3B	56% of operating cash flow.
Cash capex	Q1 2025	$77.8B	65% of operating cash flow.
Cash capex	Q2 2025	$97.3B	72% of operating cash flow.
Cash capex	Q3 2025	$105.8B	63% of operating cash flow.
Cash capex	Q4 2025	$130.7B	72% of operating cash flow.
Cash capex	Q1 2026	$148.4B	94% of cash flow. 94% of operating cash flow.

In the first quarter of 2022 they invested 56% of their operating cash flow. In the first quarter of 2026, 94%. And the data center is already financed as an asset: in March 2026, CoreWeave closed what it presents as the first investment-grade financing secured by compute infrastructure and a customer contract, for $8.5 billion. In May, closing another $3.1 billion loan, it described that kind of financing as "a scalable new asset class". CME and Silicon Data announced the first compute futures, and CME's chief executive called compute "the new oil of the 21st century".

Private AI is already technically viable. According to Epoch, in May 2026 open-weight models were about four months behind the most advanced closed ones. A 27-billion-parameter open model, Qwen 3.8, already matches August 2025's GPT-5 on Epoch's capabilities index. I have one running on a single RTX 3090.

But look at what the store of value actually is. Today what gets financed as an asset is compute plus its customer contract, and the futures CME has announced are based on GPU rental. I think the lasting value is elsewhere. Performance per dollar of new chips improves almost 50% a year, so today's GPU falls behind fast. What doesn't fall behind is land with connected power. It's the twenty-year energy contract.

Constellation is restarting Three Mile Island for Microsoft, under a twenty-year contract for about 835 MW. Amazon buys 1,920 MW of nuclear power from Talen through 2042. Google, 500 MW of small Kairos reactors. Meta, up to about 6.6 GW across Vistra, TerraPower and Oklo.

The asset is the plug. Again.

If energy isn't solved

Both futures share the same bottleneck. And if it isn't solved, there is a third path that looks marginal today.

In October 2025, Epoch estimated that a 10 GW training run spread across 23 data centers, next to gas plants with spare capacity and linked by 4,800 km of fiber, is technically feasible: at most five days of delay and under 1% extra cost. It is talking about large labs spreading their own compute over dedicated fiber, not open protocols, and a footnote says internet-based decentralized training is unlikely to produce any frontier run. I take the physical lesson: if energy doesn't come to the compute, the compute goes to the energy.

That's where decentralized AI protocols come in.

Figure 6

Decentralized training grows faster, but still trails the frontier by about 1,600x

Training compute (FLOP, log scale): frontier model record versus decentralized pretraining over the internet.

• Frontier record
• Decentralized record
• Other decentralized pretraining runs

Show the data

Series	Date	Value	Note
Frontier record	May 2020	3.1 × 10²³	GPT-3
Frontier record	Aug 2021	3.7 × 10²³	Jurassic-1-Jumbo.
Frontier record	Sep 2021	2 × 10²⁴	FLAN 137B.
Frontier record	Mar 2022	2.6 × 10²⁴	GPT-3.5 (speculative).
Frontier record	Jun 2022	2.7 × 10²⁴	Minerva.
Frontier record	Mar 2023	2.1 × 10²⁵	GPT-4
Frontier record	Dec 2023	5 × 10²⁵	Gemini 1.0 Ultra (speculative).
Frontier record	Feb 2025	3.5 × 10²⁶	Grok 3.
Frontier record	Feb 2025	3.8 × 10²⁶	GPT-4.5.
Frontier record	Jul 2025	5 × 10²⁶	Grok 4 (speculative).
Frontier record	Sep 2026	9.4 × 10²⁶	GPT-6 Astra. Epoch interval: 5 × 10²⁶ to 2 × 10²⁷.
Decentralized record	Jun 2021	8.7 × 10¹⁷	DeDLOC / sahajBERT-XL.
Decentralized record	Jul 2022	1.8 × 10¹⁹	Training Transformers Together, 38 nodes.
Decentralized record	Jan 2023	4.5 × 10²¹	SWARM, 12 nodes.
Decentralized record	Nov 2024	5.6 × 10²¹	Photon 7B, 5 nodes.
Decentralized record	Dec 2024	6 × 10²²	INTELLECT-1. Prime Intellect, 14 nodes.
Decentralized record	Jun 2025	5.8 × 10²³	Protocol Models. Pluralis Research, 8 nodes. Decentralized record in Epoch's data.
Other decentralized pretraining runs	Nov 2023	1.1 × 10²⁰	DiLoCo (Google DeepMind), 8 nodes.
Other decentralized pretraining runs	Jul 2024	3 × 10²⁰	Open-DiLoCo (Prime Intellect), 4 nodes.
Other decentralized pretraining runs	Dec 2024	8.8 × 10²¹	Nous DisTro 15B, 16 nodes.
Other decentralized pretraining runs	May 2025	1.1 × 10²¹	Templar I (Covenant).
Other decentralized pretraining runs	Jul 2025	9.6 × 10²¹	Templar II (Covenant).
Other decentralized pretraining runs	Sep 2025	3 × 10²³	Consilience 40B (Nous Research), 140 nodes.
Other decentralized pretraining runs	Oct 2025	4.7 × 10²⁰	Node0-7.5B (Pluralis), 1,642 nodes.
Other decentralized pretraining runs	Dec 2025	5.6 × 10²³	Covenant-72B. On Bittensor subnet 3, permissionless. Figure from Epoch's sheet (Dec 2025).

Jaime Sevilla, Epoch's co-founder and chief executive, estimates in an analysis under his own name that the compute of decentralized training runs over the internet has grown about 600,000-fold since 2020: about 20x a year, against 5x for the frontier. They are still far behind. I ran the numbers with Epoch's data: the largest sits about 1,600 times below GPT-6 Astra. But they are no longer toys. Covenant-72B trained a 72-billion-parameter model on about 1.1 trillion tokens over the Bittensor network, without asking anyone's permission, with at least 70 distinct participants over the run (about 17 at a time on average, each with at least eight B200s) and 94.5% compute utilization.

Sevilla thinks it's unlikely they reach the frontier this decade: he doesn't see how they would amass that much compute, and even with the same compute, the bandwidth-saving techniques would cost them performance. If energy gets solved, I agree. If it doesn't, the math changes.

A decentralized protocol doesn't need a gigawatt in one place. It needs a thousand places with a megawatt each. A solar plant's surplus at noon. A basement with contracted power nobody uses. Stray megawatts, too small for a multi-gigawatt campus, and even for the sites of hundreds of megawatts in Epoch's calculation.

If energy gets solved, scale wins, and the two futures compete over who has it. If it doesn't, AI will have to go find the megawatts wherever they are. Decentralized protocols already know how to pool compute spread across the internet. That they will learn to go after stray megawatts is my bet, not something they have shown.

That's where their spring can come from. Not from ideology. From physics.

---

If you're deciding where your AI is going to live, or you're building a data center for it, write to me at hello@alber.one. I care more about your megawatt math than your GPU list.