This Wednesday, September 2, 2026, eight AI enthusiasts and yours truly spent three hours in a meeting room, brainstorming opportunities in the AI space. What was discussed will remain confidential, but the impression we all left with is the subject of this issue.
Six months earlier, in March 2026 and before Tokenando even existed, independent analyst Benedict Evans had planted the same feeling in my mind during a fireside chat at Mobile World Congress in Barcelona, drawing on his presentation “AI Eats the World”.
Imagine asking “What does the internet mean for us?” in 1997.
We all reached the same conclusion on Wednesday, with just one caveat: this time it feels much, much faster.
I have fond memories of 1997, working at Ericsson as a fresh graduate. Everyone had a desktop at work, some connected to the internet, but none of us had email yet (an internal electronic messaging system called MEMO was installed later that year) and we shared one laptop that had to be booked and carried to site visits and client meetings. We communicated through desk phones, faxes and big printed binders.
Back then, infrastructure was expensive, business models were guesses, incumbents were very sceptical and the only people paying attention were early adopters. Bob Metcalfe, who invented Ethernet and founded 3Com, had predicted in InfoWorld in December 1995 that the internet would “catastrophically collapse” the following year; 16 months later, in April 1997, he put his column in a blender on a stage in Santa Clara and literally swallowed it. By that time in 1997, Amazon had been listed on Nasdaq for four months, Yahoo! was the most visited site on the web and Google did not even exist yet as a company.
Back to a much nearer timeline: four months ago we announced Tokenando on May 3, 2026, and 20 editions have gone out since, so this issue is rather special for us. Read on.
Front of house
The consensus of 1997 was that the front door was the business. Yahoo!, Excite, Lycos and Infoseek were portals: pages that greeted the user, sorted the web into human-curated directories and sold advertising against the traffic. Yahoo! made $67m in revenue in 1997 from around 1,700 paying advertisers, lost $23m along the way and had a market value of $2.8bn by the following March.
Netscape held the other access layer, the browser, and it charged for it: retail copies, per-seat licences to companies, fees from computer makers who shipped it. That business was shrinking fast, from 27.5 per cent of revenue in the first quarter of 1997 to about 18 per cent by the third. In January 1998 Netscape stopped charging; individuals could download Navigator and Communicator for free, and a programme called Unlimited Distribution let computer makers, internet service providers, telecoms operators and publishers redistribute the software with, in the company’s words, “no strings attached”. Revenue was supposed to come from server software and from Netcenter, which was Netscape’s own website rebuilt as a portal selling advertising, sponsorship and search placement. The original browser company answered the loss of the browser by merely copying the portals.
Both of the most valuable positions on the 1997 internet, the directory and the gateway, turned out to be places people passed through on their journey somewhere else. Actually, Tokenando looks like a portal: we track AI model prices, we publish three indices, a token price index, a compute price index and an inference spread index, we aggregate an AI news feed from 25 publisher sources, we map the landscape with 47 company profiles, a people directory of 34 names, the who is who of AI and a glossary of terms… because why not! That looks pretty much like Yahoo! in 1997, almost feature for feature, human curation included… guilty as charged!
Sold at a loss, bought with purpose
Amazon listed on Nasdaq on May 15, 1997 at $18 a share, raised $54m and closed the year with sales of $147.8m and a net loss of $27.6m, most of it from selling books below the price bookshops were barely surviving on. The market thought of it as a bookseller with bad margins. Jeff Bezos, in his first letter to shareholders, written in March 1998, described a different company: distribution capacity had gone from 50,000 to 285,000 square feet in one year, a second warehouse had opened in Delaware in November 1997, and when the choice was between margins that looked good and cash flows that compounded, “we’ll take the cash flows” said Mr Bezos. Every discounted book bought a customer, and enough customers justified the second warehouse. The loss was the entry fee to a business that needed 285,000 square feet.
Anthropic pays about $1.25bn a month to SpaceX for Colossus, the data centre near Memphis, Tennessee, and projects a 40 per cent gross margin, a figure it revised down. That is a perfectly respectable margin for a retailer but a rather poor one for a software company, which is exactly what made Amazon difficult to fathom in 1997: the numbers on the profit and loss account belonged to one kind of business but the assets being bought belonged to another. Renting capacity on that scale is the same wager Mr Bezos made on square footage, and it pays off or it does not on one question, which is whether the capacity fills.
Filed under features
In 1997 the full-text search box on Yahoo! was not their own. The company told investors that year that it supplemented its directory listings with web search results supplied by AltaVista, a division of Digital Equipment Corporation, under a non-exclusive agreement, and warned that losing that deal would hurt the business. Editors were the moat, the directory was the product and search was a convenient solution for queries the directory could not answer.
Google was incorporated on September 4, 1998 with a much better search box but no business model. Somebody else came up with how to monetise search: in February 1998 Bill Gross’s Idealab launched GoTo.com, which auctioned placement in search results and billed for it only when someone clicked. Google launched AdWords in October 2000 with 350 advertisers, sold by the 1,000 impressions, then adopted the click auction with AdWords Select in February 2002. So the best search provider on the web spent two years without a way to make money, which eventually a competitor figured out for them.
This AI wave started with chatbots. ChatGPT, Gemini and Claude are this technology cycle’s browser rather than its portal: the gateway to get anywhere, offered in a freemium model with the paid tier starting at $20 a month, travelling in the opposite direction to Netscape, which went from a paid licence to a free download. The portal model in today’s cycle belongs to the aggregators, and the two largest, OpenRouter and Hugging Face, were bought within a fortnight of each other.
Stripe agreed on August 19, 2026 to buy OpenRouter, which routes requests across more than 400 models from over 80 providers, at a price reported between $7bn and $8bn: about five times what it was worth in May, when it raised at a $1.3bn valuation. Alex Atallah and Louis Vichy founded it in early 2023, so three and a half years from start to exit. Nvidia signed on September 2, 2026 to buy Hugging Face for $12.93bn, being $11.9bn to shareholders and up to $1bn of retention equity, with completion expected in the first half of 2027 subject to regulatory approval. Hugging Face was founded in New York in 2016 as a chatbot for teenagers, moved into hosting other people’s models in 2018 and now holds more than 3m of them and 500,000 datasets for more than 18m developers, on annualised revenue The Information puts at $150m.
Neither price is explained by revenue, nor are the acquisitions’ rationales so straightforward at the outset. Stripe processes payments, Nvidia makes chips and neither of them needs a front door to sell what it sells, yet between them they have just paid more than $20bn for two. Both aggregators were charging for something, OpenRouter taking a cut on the credits developers buy and Hugging Face selling storage, hosting and enterprise features, but $50m and $150m of annualised revenue do not account for $8bn and $12.93bn. Google in 1998 had no way to charge at all. OpenRouter and Hugging Face each have one that works; what they have not had is the time to grow it into these prices, and their buyers decided not to wait.
Much, much faster
Wednesday’s three-hour brainstorm ended with a list of candidates and no clear winner, which is how it was meant to be. Over burgers and poke bowls afterwards, someone brought up USWeb, founded in December 1995 by five former Novell executives, which bought 18 web consultancies in the first 10 months of 1997 and went public that December at $7.50 a share, below the $9 to $11 it had hoped for, valuing the whole company at about $229m. It merged with CKS in 1998, was bought by Whittman-Hart, was renamed marchFIRST in March 2000 before it went bankrupt in April 2001. Being early, listed and consolidating was not the same as being right.
What has changed since 1997 is the clock. Issue Zero of The Compute, published on May 1, 2026, opened with two of my youngest son’s classmates, still in high school, who solved a WhatsApp-group problem on GPT-4.1 Nano because it did the job for less than $1 a month in token spend, against roughly 100 times that budget on a top-tier model. Less than four months later the largest monetary figure in this newsletter was $30tn, and in the final fortnight of those four months the two biggest aggregators changed hands for more than $20bn.
In 1997 nobody could tell which of the front doors, the loss-makers or the features would turn into the business; it took three years to find out. Mr Evans reckons we are standing at that same point. The difference is that we will not get three years to figure things out.
A note to our readers
Tokenando is an official media partner of AI Summit Barcelona, and we will be covering it in The Compute. The second edition runs on September 22–23, 2026 at the World Trade Center Barcelona, expecting over 10,000 attendees, 200 speakers and 50 side events across the city during AI Week.
On Wednesday, September 23, 2026 at 16:15 CEST on the Panel Stage, I will be moderating Panel 19, “From BSC to Brussels: Europe’s Supercomputing Bet on AI”. It puts public supercomputing, quantum hardware and network infrastructure in the same room: Diego Perino of the Barcelona Supercomputing Center, Raphaëlle Achach of GENCI in France, Gerald Mullally of OQC Tech and Miloš Mastník of CRA.
Readers of The Compute get 10 per cent off tickets with the code THECOMPUTE10 at aisummitbarcelona.com.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
This issue was fuelled by an office happy hour and an on the spot performance of “Total Eclipse of the Heart”, helped along by a succession of personally crafted Negroni Sbagliato and French 75. September is Negroni month, so expect more of these shenanigans in the coming issues.
Later in the week I stumbled on “Ce matin-là” by Air, which led me to The Shacks and their “Trip To Japan”. Two very different bands, 20 years apart, and both of them sound like a Sunday morning smack in the middle of the week.
References
“Amazon’s original 1997 letter to shareholders”, Amazon, reprinted from the 1997 Annual Report, March 1998. https://www.aboutamazon.com/news/company-news/amazons-original-1997-letter-to-shareholders
Amazon.com, Inc., Form 10-K for the fiscal year ended December 31, 1997, filed with the SEC on March 30, 1998. https://s2.q4cdn.com/299287126/files/doc_financials/annual/123197_10k.pdf
Netscape Communications Corp., Form 8-K, January 1998, SEC EDGAR (free client software, Unlimited Distribution programme). https://www.sec.gov/Archives/edgar/data/0000944458/000094445898000004/0000944458-98-000004.txt
Netscape Communications Corp., Form 10-Q for the quarter ended April 30, 1998, SEC EDGAR (stand-alone client revenue, Netcenter revenue). https://www.sec.gov/Archives/edgar/data/0000944458/000104746998024143/0001047469-98-024143.txt
Yahoo! Inc., Form S-3, 1997, SEC EDGAR (AltaVista supplies full-text search results under a non-exclusive agreement). https://www.sec.gov/Archives/edgar/data/0001011006/000091205797025974/0000912057-97-025974.txt
“How Yahoo! won the search wars”, Fortune, March 2, 1998. https://fortune.com/1998/03/02/how-yahoo-won-the-search-wars/
“Sage who warned of Net’s collapse eats his words”, Reuters, April 11, 1997. https://www.ibiblio.org/pjones/ils310/msg00259.html
“From the Ether: Predicting the Internet’s catastrophic collapse and ghost sites galore in 1996”, Bob Metcalfe, InfoWorld, December 4, 1995, reproduced by Elon University. https://www.elon.edu/u/imagining/expert_predictions/from-the-ether-predicting-the-internets-catastrophic-collapse-and-ghost-sites-galore-in-1996/
“Eating My Collapse Column”, Bob Metcalfe, message to the NANOG list, April 14, 1997. https://diswww.mit.edu/charon/nanog/8643
USWeb Corporation, Form S-4/A, December 1997, SEC EDGAR (initial public offering completed December 10, 1997; 18 entities acquired between January 1 and November 13, 1997). https://www.sec.gov/Archives/edgar/data/0001010724/000101287097002472/0001012870-97-002472.txt
“USWeb gets IPO off at reduced price”, Tech Monitor, December 1997. https://www.techmonitor.ai/technology/usweb_gets_ipo_off_at_reduced_price_1
“AI Eats the World”, Benedict Evans, presentation, Spring 2026 edition. https://www.ben-evans.com/presentations; recorded talk, from about minute 26.
“How we started and where we are today”, Google. https://about.google/company-info/our-story/
“Google AdWords Turns 15: A Look Back At The Origins Of A $60 Billion Business”, Search Engine Land, October 2015. https://searchengineland.com/google-adwords-turns-15-a-look-back-at-the-origins-of-a-60-billion-business-234579
“GoTo: The Forgotten Search Engine”, The History of the Web. https://thehistoryoftheweb.com/goto-forgotten-search-engine/
“Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage”, Stripe Newsroom, August 19, 2026. https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter
“OpenRouter”, Contrary Research (founding, founders, funding rounds). https://research.contrary.com/company/openrouter
“NVIDIA to Acquire Hugging Face”, Jensen Huang, NVIDIA Blog, September 3, 2026. https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
NVIDIA Corporation, Form 8-K, September 2, 2026 (purchase price and closing conditions). https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm
“What Is Hugging Face? The Open Source AI Platform”, Built In (founding, teenage chatbot origin, reported revenue). https://builtin.com/articles/what-is-hugging-face
“The line that ate the P&L”, The Compute Issue Zero, May 1, 2026.
“What a Victorian economist got right about your AI bill”, The Compute Issue I, May 7, 2026.
“Intelligence, collateralised”, The Compute Issue XVI, August 21, 2026.
“Thirty Trillion, All the Way Down”, The Compute Issue XVII, August 28, 2026.
“Anthropic Lowers Gross Margin Projection as Revenue Skyrockets”, The Information, January 22, 2026. https://www.theinformation.com/articles/anthropic-lowers-profit-margin-projection-revenue-skyrockets
AI Summit Barcelona 2026.
https://aisummitbarcelona.com/
Tokenando indices, latest settled day September 2, 2026, methodology v1.2.2; platform figures checked September 3, 2026. https://tokenando.ai/indices