AI
Agent Token Math Is Locking In a Decade of Spend
Agents already consume about 60% of AI tokens, so falling unit costs are expanding foundry, power, and Wall Street financing rather than cooling the bill.
About 60% of AI tokens now come from agents, Cisco president and chief product officer Jeetu Patel said at Goldman Sachs’s Communacopia + Technology Conference on September 8, 2026, after usage jumped 14 times since February. Fewer than 2% of people use agents as power users, and agents have not reached 2% of enterprise processes. The capex cycle is being set by machines that loop, not by a mass consumer signup.
That inversion is why falling unit prices are not cooling the build. They are making more loops cheap enough to run all day, which is the demand Nvidia, TSMC, and the large cloud buyers are still racing to feed.
Agents Already Burn Most of the Tokens
Patel’s point was not that chat got popular. It was that the mix flipped while almost nobody was using the product the way a power user does. OpenRouter’s February 2026 print was the crossover, he said, the first stretch when agents consumed more tokens than humans. Since then the agent line has gone nearly vertical.
He stacked a second split on top of the first. About 60% of the world’s compute, he said, is now spent on inference rather than training, which means the models are no longer sitting in a lab waiting for a killer app. They are already being worked. The shortage shows up before household adoption does, because each active person can stand up many agents, and those agents do not sleep.
THE TOKEN SHIFT CISCO IS TRACKING
- The mix: Agents now consume about 60% of total token volume, after a 14-times rise since February 2026.
- The user base: Fewer than 2% of humans use agents as power users, and agents have not reached 2% of enterprise process automation.
- The compute split: About 60% of global compute capacity is going to inference rather than training.
- The network tax: Cisco’s own test found an agent used 450% more network bandwidth than a person on the same task.
That last figure is why a networking company is talking like a chip vendor. Agents keep uploading skill files and memory, they retry, and they run around the clock. Patel said Cisco’s hyperscaler AI orders went from almost nothing to $9.3 billion in two years, while campus and branch networking, long stuck near 3% to 4% GDP-like growth, has been running at 20% for several quarters.
Ian Buck, Nvidia’s vice president of hyperscale and HPC computing, put a compute multiplier on the same behavior at the AI Infrastructure Summit. Agentic AI, with just four agents, is 100 times more compute-heavy than chat, he said. A person asking one question is no longer the unit of demand. A small crew of agents that plan, tool-call, and check each other is.
Why the Price Drop Makes the Bill Larger
Epoch AI’s September 22, 2026 report, “The plunging price of thought,” measured how fast a given level of model performance has gotten cheaper, not how fast a raw token sticker has moved. Luke Emberson and David Roodman found that cost has fallen about 47% per quarter since 2023, or 13 times per year, a drop they call faster than DNA sequencing, compute, lithium batteries, or electricity in the records they compared.
On January 31, 2025, they estimate, OpenAI’s o3 could hit 75% on GPQA Diamond, a PhD-level science exam, for 30 cents a question. Under 18 months later, GPT-5.6 Luna matched that score for $0.0004 a question, a 725-fold decline. The cheapness is real. So is the catch that sits next to it: reasoning models can burn far more tokens to extract that score from a smaller net, which is why a falling price per token does not automatically cut the invoice.
HOW FAST THE PRICE OF THOUGHT IS FALLING
| Measure | Epoch AI finding |
|---|---|
| Fixed performance level since 2023 | 47% cheaper per quarter, or 13× per year |
| Right after a level first hits state of the art | 66% cheaper per quarter, or 75× per year |
| Two years after that debut | 32% cheaper per quarter, or 4.7× per year |
| Pace versus DNA sequencing | 4× faster |
| Pace versus lithium batteries | 18× faster |
| GPQA Diamond at 75% (o3 to GPT-5.6 Luna) | $0.30 to $0.0004 a question, 725-fold |
A March 2025 Epoch cut, using a different method, had already shown drops of 9 times to 900 times per year across benchmarks. Gundlach and colleagues, using actual cost-to-score rather than sticker prices, land lower, at 5 times to 10 times per year. None of those paths point to a smaller industry. They point to a cheaper unit that people, and now agents, consume in bulk.
William Jevons watched the same pattern in coal and steam in 1865. Make the machine cheaper to run, and the world runs more of them. Coding is the early market where that is already visible, with developers using agents to write and check most of the code, then spending the saved hours on the next round of generation. The same pattern is what turns agent loops that multiply token bills into a planning problem for anyone who still budgets AI like a chat seat.
Unwatched loops make it worse. An agent that retries in the background for months can spend past any prompt-level forecast, and token use on the same task can swing by 2 times, sometimes much more, from run to run. Volume, including cached context that still has to move, is the bill. The unit price is the part that keeps falling.
Five Firms, $800 Billion, and a Recoup Gap
Jensen Huang, Nvidia’s chief executive, used the same Goldman Sachs conference on September 10, 2026, to repeat a forecast that still sits above most street numbers: AI infrastructure spending of $3 trillion to $4 trillion by 2030. He said the 70% revenue growth Nvidia is targeting is capped by supply, while unconstrained demand is growing more than 100%. Grace, Blackwell, and NVLink systems in 72-rack configurations were still rising 27% month over month.
Goldman Sachs strategists led by Ryan Hammond, in a September 25 note, put 2026 AI infrastructure spending by Amazon, Alphabet, Microsoft, Oracle, and Meta at $800 billion, then $1.2 trillion in 2027. Those five names are the near-term checkbook. Huang’s $3 trillion to $4 trillion is the 2030 envelope for a wider build that includes power, shells, and the rest of the stack, not a restatement of the same $800 billion.
The Hammond note also asked the question the boom’s fans like to skip. Those five firms, on Goldman’s math, need roughly $300 billion a year of AI revenue to break even on $800 billion of 2026 infrastructure spend. Payback stories from Google’s own silicon (see the next section) can be fast on a server. They do not, by themselves, close a gap that large across the whole cloud cohort.
Nvidia’s own second quarter, ended July 26, 2026, showed what the hardware layer looks like while that debate runs: $96.2 billion in revenue, up 18% from the prior quarter and 106% from a year earlier. Huang described a price path of about $18,000 per Hopper GPU system, about $25,000 for Blackwell, and about $40,000 for Vera Rubin, with some full systems at $8.5 million. Each GPU system, he said, weighs 2 tons, holds 2 million parts, and draws 250 kilowatts. Nvidia books about $50 billion of revenue per gigawatt of data center capacity per year.
Buck’s Rubin slide is the efficiency promise inside that price-up. Versus GB300 NVL72, he said Vera Rubin NVL72 can run up to 30 times faster at higher responsiveness, with up to 45 times lower cost per token, with the gap widest on agentic jobs that need quick answers. Cheaper tokens and more expensive racks are arriving together. That is only a contradiction if you still think the customer is a person typing one prompt.
Under a Year to Earn Back a TPU Server
Google Cloud chief executive Thomas Kurian used the same September conference to put a clock on Alphabet’s servers. The line is narrower than a claim about the whole capital budget, and he said so by sticking to hardware.
Our payback period on AI servers, in aggregate, is less than two years. And on our own silicon, it’s half of GPUs.
Thomas Kurian, CEO, Google Cloud, Goldman Sachs Communacopia + Technology Conference, September 2026
He added that Google’s TPU accelerator business is more than twice the size of the next-largest hyperscaler’s, and that most infrastructure contract value sits in committed five-year deals. Training price-performance, he said, is about 2.7 times the competition, with inference about 80% ahead. A server that earns back in under a year can be rolled into the next cluster while a five-year contract is still running. That is a cash-velocity argument, not proof that $800 billion of group spend has already paid off.
It also explains why custom silicon keeps spreading even while Nvidia remains the default merchant GPU. Amazon has talked about building data centers, racks, and Trainium chips in parallel so parts coming off the line can be earning in a couple of months. Every large buyer that can fund a design team now has a reason to try, because a one-year payback on in-house parts beats a two-year payback on bought ones, if the utilization shows up.
Pat Gelsinger, now a general partner at Playground Global, framed the physical task as a 10,000-times improvement across the stack: reasoning models needing 10 times more compute, future models another 10 to 100 times, and the world at most one-tenth of the way into adoption. That 10,000-times stack goal is a different quantity from Epoch’s training-FLOP path. Both point the same direction. The cheap unit is not the binding constraint. The plant is.
The Packaging Line Behind Every Accelerator
TSMC is where those plants actually get drawn. The foundry reported $40.2 billion of second-quarter 2026 revenue, up 34% from a year earlier, and held 72.5% of the global foundry market, with Samsung Foundry at 5.9%. Advanced nodes are doing the work: 3-nanometer processes were 30% of TSMC wafer revenue in the quarter, and 5-nanometer was 33%.
WHERE THE SILICON IS TIGHT
- The checkbook: TSMC lifted 2026 capital spending to $60 billion to $64 billion, with 70% to 80% aimed at advanced processes.
- The share: 72.5% of global foundry revenue in the second quarter, against 5.9% for Samsung Foundry.
- The package: Counterpoint puts the CoWoS supply-demand gap at about 20% in 2026, narrowing toward 10% in 2027 as in-house capacity heads toward 185,000 wafers a month.
- The mix: 3-nanometer wafers were 30% of TSMC wafer revenue, and 5-nanometer were 33%.
Huang listed the upstream choke points in the same breath as the demand story: packaging, DRAM, LPDDR, connectors, voltage regulators, and wafers. CoWoS-class packaging is the unglamorous gate for HBM stacks and large accelerators. TrendForce still treats TSMC’s CoWoS-L as the lead advanced package for AI through 2028, even as Intel pushes EMIB-T. A 20% packaging gap is a physical speed limit on Huang’s 70% growth target, whatever the token price is doing.
CFO Wendell Huang has been blunt that TSMC’s plan assumes the AI megatrend holds. That is a single-company concentration sitting under a global spend cycle. Hyperscalers will keep shopping Intel Foundry and Samsung as backup, because no other AI supplier is this much of a single point of failure. Backup is not the same as a second TSMC.
How Wall Street Turned GPUs Into Collateral
If the loops keep multiplying, someone has to own the building and the chips for 10 years. That someone is less and less the cloud company’s cash flow alone. Huang told the Goldman audience that Nvidia systems are now durable, rentable, and usable as loan collateral, a status he compared to ordinary infrastructure. He waved off circular-finance talk around the new platforms with a one-line return claim: put in one, get 100 back.
THE 2026 CAPITAL HOOKS
- May 18, 2026: Blackstone announces a U.S. joint venture with Google to sell data center capacity, networking, and TPUs as compute-as-a-service, with an initial $5 billion equity commitment and the first 500 MW of capacity online in 2027.
- August 10, 2026: Nvidia signs memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to set up independent compute-financing platforms meant to mobilize over $500 billion of third-party capital over time, a figure Nvidia says is not its own revenue.
- September 8 to 10, 2026: Patel, Huang, and Kurian spend the Goldman Sachs conference arguing that demand is still early, paybacks on servers can be short, and land, power, and shells are the tighter gates.
Blackstone, which manages more than $1.3 trillion and calls itself the largest global data-center landlord, named Google infrastructure veteran Benjamin Treynor Sloss as chief executive of the new TPU cloud. President and COO Jon Gray called it a chance to put capital to work at scale in AI infrastructure. The structure is the tell. Google keeps supplying TPUs, software, and services. Blackstone supplies the equity and the energy-and-real-estate machine. Customers get a second door into TPU capacity besides Google Cloud.
The $500 billion Nvidia platforms are the merchant-GPU version of the same idea. Apollo, BlackRock, Brookfield, Goldman Sachs, and KKR join Blackstone in pooling long-duration money so labs and clouds can rent factories instead of buying every rack on the balance sheet. Huang’s line that chips have become an investable asset class is a financing claim. It only stays true if those racks keep earning, which runs back to Patel’s 60% agent mix and Kurian’s sub-two-year server clock.
Land and Power Now Set the Pace
Huang said Nvidia is more worried about data-center land, power, and shell than about parts, because the company’s supplier web can still move wafers and boards. Australia is the example he put on the table: a 2 gigawatt commitment for 2027 that he pegged at $80 billion of infrastructure, or about $40 billion per gigawatt of build, against the $50 billion of revenue Nvidia says a gigawatt of capacity can generate in a year. Those are different meters. One is what it costs to stand the plant up. The other is what Nvidia books if the plant is full.
Epoch’s longer training path still assumes the plants get built. Frontier training compute has been rising about 4 times to 5 times a year, and the group’s scaling paper still treats 2e29 FLOP training runs by 2030 as likely feasible, a 10,000-times jump from GPT-4’s roughly 2e25 FLOP. Power for those runs is already a gigawatt-class problem. Latency then pushes clusters to sit close together, which is why low-rise training yards start looking like high-rises in the slideware.
Sarah Friar, OpenAI’s chief financial officer, has argued that even a freeze in new model work would leave a huge stock of unused intelligence to productize. Dario Amodei, Anthropic’s chief executive, has asked labs to pace frontier releases so controls can keep up, and has noted that those controls mean more training, not less. Neither comment is a capex holiday. Safeguards and productization both consume clusters.
The next few years can still include pauses, consolidations, and ugly recoup quarters. Custom chips will keep taking jobs Nvidia used to own by default, and after 2030 that pressure gets sharper. None of that undoes the September print from Cisco. Agents already take about 60% of the tokens, almost no one is a power user yet, and the price of a correct answer is still falling 13 times a year. The factories, the substations, and the funds are what have to catch that curve.
Disclaimer: This article is news reporting and analysis of company remarks, research notes, and industry data, and it is informational only. It is not investment advice, a recommendation to buy or sell any security, or a forecast of returns on AI infrastructure, chips, or cloud stocks. Readers should consult a licensed financial adviser before acting on capital-spending, valuation, or payback figures. The numbers and project statuses here reflect the named companies and research groups on the dates given in the piece and can change with new earnings, contracts, or supply updates.
-
AI5 months agoFable 5 Came Back Under a Commerce On-Off Switch
-
AI5 months agoGoogle’s SpaceX GPU Lease Has a Sept. 30 Deadline
-
CRYPTO5 months agoPlasma One’s XPL Locks Face a 1.81 Billion Cliff
-
APPS5 months agoDGO’s Rs 549 World Cup Pass Cost Fans Sleep and Data
-
AI5 months agoMoonshot AI’s $30 Billion Ask Became a $35 Billion Close
-
NEWS6 months agoColorOS 17 Device List Spans Oppo, OnePlus and Realme
-
GAMING5 months agoThe RTX 4050 Under Rs 70,000 Hides a Wattage Gap
-
AI5 months agoBoris Cherny’s Loops Still Demand a Higher Production Bar
