AI Got 80% Cheaper. Why Does Everything AI Touches Cost More?
The price of AI per token keeps falling, yet AI's growth is pushing up the price of phones, PCs, memory and electricity. The data shows both are true at once, and explains why.
Source: Alphabet Q3 2026 earnings release. Values in billions of US dollars.
On 30 July 2026, OpenAI cut the price of one of its AI models by 80%. Price cuts like this have become routine: running a capable AI model costs a small fraction of what it did two years ago.
Yet over the same period, phones, PCs, memory chips and, in parts of the US, electricity have become more expensive, and AI is a big part of the reason.
So is AI making technology cheaper or more expensive? The data says both, depending on whether you look at the price of one unit or the total bill.
AI is getting much cheaper per unit
AI models are priced per token, a small chunk of text, roughly three-quarters of a word. Developers pay a set price per million tokens they send to a model and per million tokens it writes back.
On 30 July 2026, OpenAI announced price cuts for two of its GPT-5.6 models, three weeks after launching them (OpenAI announcement):
| Model | Before (input / output, per million tokens) | After | Cut |
|---|---|---|---|
| GPT-5.6 Luna | $1.00 / $6.00 | $0.20 / $1.20 | 80% |
| GPT-5.6 Terra | $2.50 / $15.00 | $2.00 / $12.00 | 20% |
Source: OpenAI developer community announcement, 30 July 2026. OpenAI’s most capable model in the family, Sol, was not part of this cut.
OpenAI said the cuts came from making its systems more efficient, including about 20% lower serving costs, and that “these improvements are being passed on to everyone using the API, Codex, and ChatGPT.”
This is part of a much longer trend. Epoch AI, which tracks AI costs, measured how quickly the price of reaching a fixed level of performance falls. For GPT-4-level performance on a test of PhD-level science questions, the price fell about 40 times per year. Across different tests and performance levels, the declines ranged from 9 to 900 times per year (Epoch AI, March 2025).
Epoch AI measures the price of reaching a fixed level of performance over time; the 40× per year figure equals a 97.5% fall in one year. Epoch notes the fastest declines are recent and may not persist.
If you only looked at the price tag, you would expect spending on AI to be falling fast.
But AI use is growing even faster
Prices fell, and usage exploded. Google, which publishes its own token counts, gave this timeline at its I/O conference in May 2026 (Google):
Self-reported by Google. 3.2 quadrillion = 3,200 trillion. Source: Google, I/O 2026 keynote, 19 May 2026.
The growth has not slowed since. On 28 October 2026, Google said its AI services for developers were processing “nearly 32 billion tokens per minute,” 40% more than the previous quarter and “more than four times what we processed a year ago” (Google).
It is not just Google. OpenRouter, a service that lets developers reach many AI models through one connection, publishes its own traffic. Its weekly volume rose from about 0.5 trillion tokens in mid-January 2025 to 128.4 trillion in mid-September 2026, “roughly a 250x increase in 20 months” (OpenRouter, 22 September 2026). OpenRouter also estimates that since February 2026, AI agents (software that runs tasks on its own) have used more tokens than people typing to chatbots.
Price × volume: why the total bill keeps rising
This is the same arithmetic as the 1 TB Test, turned the other way round. Total spending equals the price of each unit times the number of units:
A simple example using the figures above: if the price per token falls 80% (to 0.2 times) while usage grows 7 times, as Google’s did between May 2025 and May 2026, the total bill becomes 0.2 × 7 = 1.4 times larger. Every token is cheaper, and the bill is still 40% higher.
Economists have a name for this pattern. In 1865, William Stanley Jevons observed that more efficient steam engines did not reduce Britain’s coal use; they increased it, because cheaper power found many more uses. AI is following the same path: as each answer gets cheaper, people and businesses use far more of them.
The companies building AI capacity show what that means in dollars. Alphabet’s spending on property and equipment, mostly data centers, servers and chips, more than doubled in a year:
| Alphabet, quarter ending September | 2025 | 2026 | Change |
|---|---|---|---|
| Purchases of property and equipment | $23.95 billion | $51.36 billion | +114% |
| Google Cloud revenue | $15.16 billion | $30.71 billion | +103% |
Source: Alphabet Q3 2026 earnings release. Google Cloud revenue includes AI infrastructure and services along with other cloud products.
Microsoft reported the same direction: $45.2 billion of capital spending in the quarter to September, up about 30% from a year earlier, with Azure cloud revenue up 48% (CNBC, 28 October 2026). Across the largest cloud providers, research firm TrendForce projects capital spending will nearly double (up 98%) in 2026 (TrendForce, 25 August 2026).
Where the cost lands: everything AI touches
All of that spending buys physical things: memory chips, processors and electricity. AI data centers are competing for the same supplies as phones, PCs and households, and that is where higher prices show up for everyone else.
Memory
Memory has been hit hardest. TrendForce projects that server memory contract prices will rise about 270% in 2026, and that memory will make up 47% of the largest cloud providers’ capital spending in 2026 and 68% in 2027 (TrendForce). Microsoft’s finance chief said its spending would accelerate “inclusive of higher memory and storage costs” (CNBC).
The same chips go into consumer devices, so their prices have risen too:
Each figure covers a different product and period; see sources. The DDR5 figure compares the cheapest tracked 32GB DDR5 kit before the shortage (about $95) with 16 October 2026 ($456). Source: Capital and Compute; Counterpoint Research, 3 Nov 2026; Gartner, 12 Oct 2026.
Sources: Capital and Compute – DDR5 price tracker; Gartner (“DRAM costs surged five times year-over-year in the second quarter”); Counterpoint.
Higher prices have reduced how many devices people buy. Worldwide PC shipments fell 22.4% and smartphone shipments fell about 20% in Q3 2026, as we covered in PC and phone shipments fell about 20%. Did anyone actually spend less? Microsoft’s device and Windows sales fell 22% in the quarter, which CNBC linked partly to higher memory and storage costs raising device prices, alongside the end of Windows 10 support and weak PC demand (CNBC).
Electricity
Data centers also need power, and lots of it. PJM runs the largest power grid in the US, covering 13 states and Washington, D.C. Each year it holds an auction for “capacity”, payments that guarantee enough power plants will be available at peak times. Those costs flow into electricity bills.
2027/28 and 2028/29 cleared at the price caps in effect for those auctions. Values for most of PJM's footprint; some zones cleared higher. Source: Utility Dive (citing PJM), 18 Dec 2025 and Jul 2024; RTO Insider, 14 Jul 2026.
The price rose from $28.92 per megawatt-day for 2024/25 to $269.92 for 2025/26 (Utility Dive), then to $329.17 and $333.44 for the next two years, the second of those at the region-wide price cap (Utility Dive). The July 2026 auction for 2028/29 again cleared at its cap, $325, and came up 6,831 megawatts short of PJM’s reliability target (RTO Insider).
Data centers are not the only reason, but PJM has said the recent rise in its demand forecast was “almost entirely driven by data centers” (Utility Dive).
So, is AI making tech cheaper or more expensive?
The answer depends on who you are:
| If you are… | What has happened | Net effect |
|---|---|---|
| A developer or business using AI models | Price per token keeps falling; the same task costs far less than a year ago | Cheaper per task, though total AI bills often rise as usage grows |
| A company building AI capacity | Spending more than doubling, with memory a growing share | Much more expensive |
| A consumer buying a phone, PC or console | Memory-driven price rises; fewer low-cost models | More expensive |
| An electricity customer on the PJM grid | Capacity prices up about 11 times since 2024/25 | More expensive |
All four are true at once. AI makes each unit of intelligence cheaper, and that cheapness drives so much more use that the total demand for chips, memory and power goes up, and everyone who shares those supplies pays more.
None of this means falling AI prices are a trick, or that any company is doing something wrong. Lower prices reflect real efficiency gains, and many users genuinely save money per task. The point is simply that a lower unit price and a higher total cost can come from the same trend.
What to watch next
- Usage vs prices: if usage growth slows while prices keep falling, total AI spending would start to level off. Google’s 40% quarter-on-quarter growth in API tokens shows no sign of that yet.
- New supply: TrendForce expects new memory factories ramping up in the second half of 2027 to add about 27% to server memory supply that year (TrendForce). Until then, AI and consumers will keep competing for the same chips.
- Power: PJM’s next capacity auctions will show whether the cap keeps prices from rising further.
For the wider story of the memory shortage, see Why RAM, SSD and Server Prices Exploded in 2025–2026.
Key terms
- Token: a small piece of text an AI model reads or writes, roughly three-quarters of an English word. AI services are priced per million tokens.
- Inference: running a trained AI model to answer a question or complete a task.
- Capital spending (capex): money spent on long-lasting assets such as data centers, servers and chips. Alphabet reports it as “purchases of property and equipment”.
- Capacity auction: an auction where a power grid pays power plants to be available at peak times. The cost is passed on through electricity bills.
- Jevons paradox: when making something more efficient makes it so much more useful that total use, and total cost, goes up.
Notes on the data
- Token counts are self-reported by Google and OpenRouter and are not independently audited. OpenRouter’s figures cover traffic through its own service only.
- Epoch AI’s price trend measures the cost of reaching a fixed level of performance; it is not the price of any single model, and the fastest declines may not continue.
- The 0.2 × 7 example combines a price cut and a usage figure from different sources to illustrate the arithmetic. It is not a measurement of any company’s actual AI bill.
- Alphabet’s capital spending covers all its infrastructure, not only AI, though Alphabet attributes much of the growth to AI demand.
- PJM capacity prices are for most of the PJM region; some zones cleared at different prices. How much of each electricity bill comes from capacity costs varies by state and utility.
Sources
AI prices: OpenAI – GPT-5.6 price drop announcement (30 Jul 2026) · Epoch AI – LLM inference price trends (Mar 2025)
AI usage: Google – I/O 2026 keynote (19 May 2026) · Google – Q3 2026 earnings remarks (28 Oct 2026) · OpenRouter – The Token State of Play (22 Sep 2026)
Spending: Alphabet – Q3 2026 earnings release · CNBC – Microsoft Q1 FY27 earnings (28 Oct 2026) · TrendForce – memory and cloud provider capex (25 Aug 2026)
Consumer prices: Capital and Compute – DDR5 price tracker · Counterpoint – Q3 2026 smartphone revenue (3 Nov 2026) · Gartner – Q3 2026 PC shipments (12 Oct 2026)
Electricity: Utility Dive – PJM 2025/26 auction (Jul 2024) · Utility Dive – PJM 2027/28 auction (18 Dec 2025) · RTO Insider – PJM 2028/29 auction (14 Jul 2026)