SuperworthIntelligence keeps getting cheaper, and the AI bill keeps growing
Each answer costs less every year, and buyers spend more in total. Seats and hours are priced for the loss. The chip and memory makers are priced as if the spending ends.
A thesis by Superworth
Helped
- NVDASells the compute. Fwd PE 18.9, cheaper than 98.7% of its ten years, with consensus adjusted EPS up 71.2%.
- AVGOCustom AI chips and networking. Fwd PE 20.5 against a ten-year median of 15.0.
- MUMemory, the least obvious winner. Fwd PE 7.2 near its 9.5 median, on trailing revenue up 167%.
- ANETData center networking. Fwd PE 43.4, higher than 93.8% of its ten years. Already priced for more spending.
- ETNPower equipment for data centers. Fwd PE 28.8, higher than 87.9% of its ten years.
- VRTCooling and power for data centers. Fwd PE 31.5 against a 27.1 median, 34% below its May high.
- AMZNAWS grew 37% and Amazon holds a stake in Anthropic. Fwd PE 26.4, cheaper than 99.3% of its ten years.
- MSFTAzure grew 43%. Fwd PE 25.9 against a three-year median of 31.3.
- GOOGGemini, its own chips and a cloud that grew 82%. Fwd PE 25.0, above its three-year median of 21.7.
Exposed
- ADBECreative software sold by the seat. Fwd PE 8.7, cheaper than 98.9% of its ten years, revenue up 12.0%.
- NOWWorkflow software sold by the seat. Fwd PE 28.7 against a ten-year median of 71.4, revenue up 22.2%.
- ACNBills clients for people's time. Fwd PE 12.4, cheaper than 97.9% of its ten years. Reports on 1 October.
- CTSHIT outsourcing. Fwd PE 9.4 against a ten-year median of 16.5.
Each answer costs less, and the total bill keeps rising
Epoch AI found the price of one fixed level of AI on a science test fell 40x a year. Use grew faster: Google handled 3.2 quadrillion tokens a month in May, up from 480 trillion a year before. The four biggest spenders put 510.6B into capex over the last four quarters, up from 291.5B a year earlier. Alphabet alone went to 132.4B from 67.0B. Some heavy users now pay for two top-tier plans instead of one. Cheaper intelligence, more of it bought.
Software sold by the seat is priced for cheaper agents
When an agent does a seat's work for cents of compute, the seat loses pricing power. Adobe trades at 8.7 times the next twelve months, cheaper than 98.9% of its ten years (median 30.3). ServiceNow is at 28.7 against a median of 71.4. Revenue still grew 12.0% and 22.2% over the year. The market already prices the loss. Forward PE history is rebuilt with hindsight.
The sellers of hours sit near ten-year lows
If a model does an hour's work for less each year, clients buy the output, not the hours. Accenture trades at 12.4 times the next twelve months, cheaper than 97.9% of its ten years (median 25.1). Cognizant is at 9.4 against 16.5. Both rose over 40% from their June lows, and Accenture reports on 1 October. Cheaper intelligence could also bring more work installing it. Forward PE history is rebuilt with hindsight.
The clouds sell the tokens and are not priced at a premium
Anthropic told investors its revenue run rate hit 65B in July, against about 10B of revenue in all of 2025. Last quarter AWS grew 37%, Azure 43% and Google Cloud 82%. Yet Amazon trades at 26.4 times the next twelve months against a three-year median of 33.2, and Microsoft at 25.9 against 31.3. Alphabet, at 25.0 against 21.7, is the one priced up. Forward PE history is rebuilt with hindsight.
The compute seller is priced as if the spending stops
Nvidia trades at 18.9 times the next twelve months, cheaper than 98.7% of its ten years (median 33.5), while consensus sees adjusted EPS up 71.2% and revenue up 77.4%. Broadcom, which designs custom AI chips, is at 20.5 against its 15.0 median. A low multiple on fast growth says the market doubts the spending lasts. The thesis says cheaper tokens make it last. Nvidia reports on 18 November. Forward PE history is rebuilt with hindsight.
The plumbing is where the idea is already in the price
Arista, which connects the chips, trades at 43.4 times the next twelve months, higher than 93.8% of its ten years (median 30.7), on adjusted EPS seen up 35.0%. Eaton, which sells power gear, is at 28.8, higher than 87.9% of its ten years. Vertiv, which cools and powers data centers, is at 31.5 against a 27.1 median. Here the market already pays for more spending. Forward PE history is rebuilt with hindsight.
Memory is the surprise, and it is priced as a cycle
Longer context and more agents need more memory per answer. Amazon raised its 2026 capex plan to 220B, citing memory prices. Micron's trailing revenue rose from 16.2B in late 2023 to 90.3B. Yet it trades at 7.2 times the next twelve months, below its 9.5 median, a multiple for a peak. On 30 September it reported 54.2B for the August quarter and guided to 61.5B, above the 57.2B consensus. Forward PE history is rebuilt with hindsight.
What would prove it wrong: prices fall faster than use grows
If each token's price falls faster than use grows, the bill shrinks. In one September week Opus 5.5 launched 40% cheaper to run and GPT-6 Sol at half the price, as cheaper open models compete. Oracle shows the overbuild risk: trailing free cash flow went from 12.3B in early 2024 to minus 28.7B as capex reached 75.7B, and the stock is 58% below its high. Watch Oracle on 9 December.
What has happened lately
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasksThe price to reach a fixed benchmark score fell between 9x and 900x a year depending on the task, and 40x a year for GPT-4's level on PhD-level science questions.
- Google, Google I/O 2026: Sundar Pichai's opening keynoteGoogle processes over 3.2 quadrillion tokens a month, 7x a year earlier, up from 480 trillion at I/O 2025 and 9.7 trillion two years ago.
- Google, Alphabet earnings call Q2 2026: Sundar Pichai remarksCloud revenue grew 82% and backlog reached 514B, while the cost of an AI Mode response fell to its lowest since launch. Cheaper per answer, more answers sold.
- Microsoft, Microsoft Cloud and AI strength fuels fourth quarter resultsAzure and other cloud services revenue grew 43%. Microsoft frames the business as the 'cost-to-outcome curve': helping customers turn tokens into results.
- CNBC, Amazon hikes 2026 capex to $220 billion due to higher memory costsAWS grew 37%, its fastest since 2021, and Amazon raised its capex plan from 200B to 220B, citing memory prices, while saying it still lacks capacity to meet demand.
- CNBC, Anthropic tells investors annualized revenue run rate climbed to $65 billion in JulyA model seller's revenue run rate rose about sevenfold in a year to 65B, against roughly 10B of revenue in all of 2025, while its price per token kept falling.
- SEC, Micron fiscal fourth quarter 2026 resultsRevenue of 54.23B for the August quarter and 133.19B for the year, with guidance of 61.5B for the next quarter. Reported after the numbers in this thesis were read.
- CNBC, Anthropic and OpenAI roll out cheaper models in first release since call for slowdownThe counter-case in one week: Opus 5.5 costs about 40% less to run than Opus 5 and GPT-6 Sol halves API prices, as both labs face cheaper open-weight rivals.
- Oracle, Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure RevenuesCloud infrastructure revenue rose 121% and contracted backlog reached 664B, but free cash flow was minus 5B in the quarter as Oracle kept building.
- CNBC, Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. historyThe counter-case moved prices once before: a cheap, efficient open model from DeepSeek sent Nvidia down 17% in a day, after DeepSeek said it built its model for under 6M on lesser chips.
Headlines link to the original source.
Brought to you by Superworth
An idea, tested against what the prices already assume.
A thesis tests an idea against prices. It is not investment advice.