The AI infrastructure race now has a veteran voice calling it a rerun. In an essay headlined The AI infrastructure race: I’ve seen this movie before, Fast Company argued on October 24, 2026 that technology companies are running into infrastructure built for a different era, and that the firms building artificial intelligence are starting to turn into infrastructure companies themselves.
The author says their own focus moved from optics for autonomous vehicles to optical networking for AI infrastructure. That path matters because light, not silicon alone, now decides how fast a training cluster can talk to itself. We covered the manufacturing side of that bet in our look at Intel’s 300mm indium phosphide photonics claim.
This analysis separates what Fast Company confirms from what it leaves open, then lays the story against the model launches, spending signals and bottlenecks that surround it. Where the public record is thin, we say so instead of filling the gap with guesses.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Fast Company Piece Confirms About the AI Infrastructure Race
The confirmed claims are short. Fast Company says technology companies face infrastructure designed for another era. It says companies building AI are beginning to become infrastructure companies. And it says some of them are taking infrastructure problems into their own hands instead of waiting for traditional power and telecommunications providers.
That is a thesis, not a ledger. The excerpt we could verify names no products, no partner companies, no dollar figures and no on-record quotations. We therefore treat the essay as a framing device and anchor every number below to a separate named source.
The framing still carries weight. For three years the AI conversation centered on models and the chips that train them. The Fast Company argument moves the spotlight to the layers underneath: the fiber, the switches, the substations and the cooling loops that decide whether a model gets trained at all.
From Autonomous-Vehicle Optics to Optical Networking for AI
The author’s career turn is a small story with a large signal. Optics for autonomous vehicles and optical networking for AI both depend on controlling light with precision and manufacturing it at volume. Engineers who spent years on one problem can carry much of that know-how into the other.
The demand curves look different, though. Self-driving programs have stretched timelines again and again. Tesla, for example, has pushed unsupervised FSD toward Q2 2027 and confined it to robotaxi zones, as we reported in our piece on the unsupervised FSD delay. AI clusters, by contrast, are being ordered and installed now.
That contrast explains why optical talent is migrating. A lidar or camera supplier waits on regulators and road tests. An optical networking supplier waits on a purchase order from a company that already has the GPUs and needs a way to connect them.
Why Optical Networking Sits at the Center of AI Clusters
AI workloads push enormous volumes of data between accelerators, memory and switches. Copper hits physical limits as speeds climb, so the industry keeps moving optical components closer to the switching silicon. Co-packaged optics, which places the optical engine next to the switch chip instead of in a pluggable module, is the clearest expression of that move.
We could not verify a current, sourced figure for bandwidth demand or for the optical market’s size, so we skip both. What we can do is show how the largest supplier talks about the problem. NVIDIA founder and CEO Jensen Huang said in the company’s silicon-photonics switch announcement: “AI factories are a new class of data centers with extreme scale, and networking infrastructure must be reinvented to keep pace.” In the same release he said: “By integrating silicon photonics directly into switches, NVIDIA is shattering the old limitations of hyperscale and enterprise networks and opening the gate to million-GPU AI factories.”
Treat those lines as a vendor’s pitch. They still show where the money is heading. When the company selling the GPUs says the network must be reinvented, networking has moved from a supporting cost to a strategic constraint.
What Huang Says About Photonics Capacity
Huang pressed the same point in a May 2026 interview. Speaking to CNBC on May 29, he said: “When you look upstream, you come to the conclusion that we’re starting to scale our silicon photonics technology.” He added: “Which means the amount of silicon photonics technology capacity that we need is substantially higher than the world has today.”
A shortage claim from a seller deserves some skepticism. It also fits the Fast Company thesis neatly. If the world lacks photonics capacity, the companies that depend on it have a reason to secure supply themselves, which is what becoming an infrastructure company looks like in practice.
Nine Models in Five Weeks: The Demand Behind the Buildout
Infrastructure gets built because someone expects to use it. The model calendar shows that expectation clearly. One release tracker counted 31 model releases in September 2026, and another listed more than 40 releases from 23 providers in the same month.
The table below collects the headline launches between September 3 and October 7, using prices as reported by release trackers. Prices change quickly, so check each provider’s page before budgeting. For performance rankings of the same models, see our LLM leaderboard breakdown.
| Date (2026) | Model | Provider | Reported price per 1M tokens (input / output) | Detail |
|---|---|---|---|---|
| Sept. 3 | GPT-6 Astra | OpenAI | $10 / $50 | Surcharge above 272,000-token prompts |
| Sept. 22 | GPT-6 Sol | OpenAI | $2 / $10 | Price as reported at launch |
| Sept. 22 | GPT-6 Luna | OpenAI | $0.10 / $0.50 | Price as reported at launch |
| Sept. 22 | Claude Opus 5.5 | Anthropic | $4 / $20 | Price as reported at launch |
| Sept. 28 | Claude Sonnet 5.5 | Anthropic | $2 / $10 | Price as reported at launch |
| Sept. 29 | GPT-6.1 Sol | OpenAI | $2 / $10 | Cached input cut to $0.10 |
| Sept. 30 | Gemini 4 Argon | Not confirmed | Cybersecurity focus, 1M-token context | |
| Oct. 6 | Mistral Large 4 (public preview) | Mistral | $0.68 / $2.09 (launch promotion) | 1.05 trillion parameters |
| Oct. 7 | Claude Haiku 5.5 | Anthropic | $0.10 / $0.50 | Up to 100,000-token prompts in one tracker |
Nine launches in roughly five weeks means nine sets of weights to serve, test and scale. Every one of them lands on the same constrained pool of accelerators, memory and network capacity.
Cheaper Tokens, Bigger Context Windows, Heavier Load
The pricing pattern in the table points one direction. Entry tiers such as GPT-6 Luna and Claude Haiku 5.5 sit at $0.10 per million input tokens, and GPT-6.1 Sol dropped its cached-input rate to the same figure. Meanwhile Gemini 4 Argon lists a one-million-token context window.
Lower unit prices rarely shrink total spending. They invite more calls, longer prompts and more agents running in loops. A million-token context window also means far more memory traffic per request, which lands directly on the networking and HBM layers the Fast Company essay says companies must now secure.
That is the quiet link between a price list and a power contract. When a provider cuts a token price, it is betting on volume, and volume needs racks, fiber and electricity that someone has to own or lease. Our comparison of cache pricing across Gemini, ChatGPT and Claude APIs covers how the three labs price cached input.
Power and Telecom: Companies Stop Waiting on Utilities
The most pointed line in the Fast Company piece concerns power and telecommunications. Some companies, it says, are solving those problems themselves instead of waiting for traditional providers. The excerpt gives no project names, so we cannot say who or how much.
We can point to a documented example on the telecom side. On October 21, 2026, NVIDIA and Nokia used the opening day of GTC Berlin to say that nearly 20 operators worldwide have announced AI-RAN collaborations built on Nokia’s anyRAN software and NVIDIA’s Aerial RAN Computer, according to the companies. NVIDIA also said T-Mobile completed the first AI-RAN call on a commercial network. Our full report covers the NVIDIA and Nokia AI-RAN push.
That story runs in the opposite direction from the Fast Company thesis in one respect. Here a chip company moves into the telecom network instead of waiting for carriers to modernize. It is the same instinct, just aimed at cell towers.
On power, the available research gave us no reliable 2026 figures for data-center electricity demand, grid-connection queues or regional shortages. We will not invent them. The direction is still plain: availability of electricity has become a strategic limit on expansion, and firms that can secure their own supply will move faster than those that cannot.
The Capex Signal: Meta Raises, Tesla Cuts
Spending decisions reveal how much each company believes in the buildout. On October 22, 2026, Meta and Tesla reported third-quarter results that pulled in opposite directions, according to Investor’s Business Daily. Meta raised its AI capital spending outlook to $145 billion, while Tesla trimmed its own. Futures slid that morning, with Nasdaq-100 contracts down 0.7% and S&P 500 contracts down 0.4%.
Our write-up of the Meta and Tesla capex split shows the market rewarding the raise and punishing the cut. That reaction matters for the Fast Company argument. If investors reward firms that spend more on infrastructure, more AI companies will follow Meta’s path and own the plumbing instead of renting it.
We did not find verified 2026 capex totals for Microsoft, Google or Amazon in this research pass, so we leave them out. Readers who need those figures should pull them from each company’s filings.
Memory, Chips and Cooling: The Other Bottlenecks
Optical links are one constraint among several. The table below maps the pressure points we have documented in recent coverage, with the named source behind each claim.
| Layer | What named sources report | Source |
|---|---|---|
| Optical networking | Intel says it can supply 300mm indium phosphide silicon photonics at scale as co-packaged optics enters AI data center budgets | Intel at OCP Global Summit 2026 |
| Telecom edge | Nearly 20 operators announced AI-RAN collaborations on Oct. 21 | NVIDIA and Nokia, GTC Berlin |
| Software and chips | Google open-sourced TorchTPU on Oct. 20-21 to run PyTorch code on TPUs | Google, PyTorch Conference |
| Memory | HBM and advanced DRAM treated as a central AI constraint | Seeking Alpha report, Oct. 22 |
| Cooling and water | Data centers used 222 billion liters for cooling in 2025, possibly 644 billion by 2030 without adaptive measures | Rystad Energy estimates |
| Capital | Meta raised AI capex to $145 billion, Tesla cut its outlook | Investor’s Business Daily, Oct. 22 |
| Funding | Tenstorrent raised a $1.5 billion Series E at a $6.2 billion valuation | Crunchbase News |
The water row deserves attention. Rystad’s estimate of 222 billion liters in 2025 and a possible 644 billion by 2030 turns cooling into a siting and permitting problem, not just an engineering one. We examined it in our report on AI cooling and water use.
Competitive Comparison: Four Ways to Own the Stack
If AI companies are becoming infrastructure companies, they are choosing different layers to own. Four strategies stand out in the past week’s news.
Own the Silicon and the Software
Google’s TorchTPU release lets PyTorch developers point existing code at TPUs, per Google’s announcement at the PyTorch Conference in San Jose. Our analysis of TorchTPU and Nvidia’s lock-in notes that the lock-in framing came from a trade outlet, not from Google. The strategic point stands either way: control of the chip and the developer path lowers dependence on any one supplier.
Own the Network and the Edge
NVIDIA pursues this route twice. Its photonics switches target the data center network, and its AI-RAN work with Nokia targets the carrier network. Intel’s manufacturing claim pursues the third angle, supplying the optical parts themselves. In all three cases the company moves beyond selling a processor and into the connective tissue between processors.
The remaining two strategies are financial. Meta is lifting its own capital spending. Chip startups such as Tenstorrent raise venture rounds, in its case $1.5 billion at a $6.2 billion valuation, to build alternatives. A company can buy its way into infrastructure, or it can invest in whoever builds it.
Historical Context: Why “I’ve Seen This Movie Before” Resonates
The Fast Company headline invites a historical reading, though the excerpt does not spell out which earlier film the author means. The obvious candidate is the late-1990s telecom buildout. Carriers laid fiber faster than traffic arrived, and the early-2000s bust left much of that capacity unused for years. The surviving networks later carried the streaming and cloud boom, but many of the original builders did not.
Others use the same phrase for a darker point. A news.com.au headline quotes Big Short investor Steve Eisman saying “I’ve seen this movie before” while he warns about the AI boom. We cite only the headline here and make no claim about his reasoning.
The two readings pull in different directions. One says infrastructure outlasts the companies that overbuild it. The other says the overbuilders get punished. Both can be true at once, and the 19th-century railroads and the 1990s fiber networks both fit that pattern.
What looks different this time is who holds the risk. In the fiber era, specialist carriers carried the debt. Today the spenders include some of the largest and most profitable companies on earth, which softens the downside for them while raising the stakes for suppliers and lenders further down the chain.
Market Impact: Who Gains and Who Carries the Risk
Suppliers of the scarce layers gain first. Optical component makers, memory vendors and cooling specialists sell into a market where buyers cannot easily wait. Huang’s remark that photonics capacity must grow well past today’s levels, if accurate, would describe a seller’s market for years.
Model labs face a harder equation. Token prices fell across the nine launches above while the hardware beneath them stayed scarce. Labs that own or lock in capacity can hold margins, and those that rent at spot rates absorb the squeeze. This explains why Fast Company sees AI developers morphing into infrastructure operators.
Smaller teams should watch for second-order effects. Cloud availability, quota limits and regional pricing will increasingly follow where power and network capacity exist. Anyone building on a frontier API should plan for price swings and capacity limits, and keep a fallback model ready.
For more coverage of model launches and the spending behind them, browse our AI and machine learning section.
Quote Check: Reading Vendor Statements Carefully
NVIDIA’s own words frame the shift as a reinvention of computing. In an announcement of its strategic partnership with Coherent, Coherent’s press release quotes Huang: “Computing has fundamentally changed. In the age of AI, software runs on intelligence with tokens generated in real time by AI factories for every interaction and every context.”
The phrase “AI factories” does a lot of work in these statements. A factory implies owned equipment, steady supply lines and a power bill, which is exactly the vocabulary of infrastructure. Whether or not the label sticks, it signals how NVIDIA wants customers to think about their purchases.
Every on-record statement we found comes from one executive at one company. That narrows the evidence base. We have no verified quote from the Fast Company author, and we found no independent engineering benchmark for photonics deployments. Both gaps matter when weighing the claims.
Four Predictions for the Next 12 Months
These are our own expectations, not reported facts. We built them from the sources cited above.
- More direct deals for power and fiber. Expect additional AI developers to sign their own energy and connectivity arrangements instead of relying on standard utility queues, extending the pattern Fast Company describes.
- Optical networking moves into earnings calls. Co-packaged optics and photonics supply will show up as named line items in supplier and hyperscaler commentary, not just in conference keynotes.
- Token prices keep falling while total compute spend rises. The September and October launches already show entry tiers at $0.10 per million input tokens, and we expect competitors to match or undercut them.
- Investors split spenders by visible AI revenue. The Meta and Tesla reaction on October 22 suggests the market will reward capex tied to clear AI demand and punish capex without it.
What We Could Not Verify
Several claims circulating around this story lack a confirmed source, and we have kept them out of the analysis. The Fast Company excerpt names no companies beyond the publisher, no products, no numbers and no dates besides October 24, 2026. Our search found no verified co-packaged optics deployment percentages, no sourced optical market size, and no reliable 2026 data-center power figures.
The Gemini 4 Argon price is also unconfirmed in our research, and trackers disagree on some launch details, so we used the more conservative readings. If you cite the model table, check the provider pages first. We will update this piece as named sources publish firmer data.
FAQ: The AI Infrastructure Race
What did Fast Company report about the AI infrastructure race?
On October 24, 2026, Fast Company argued that technology companies face infrastructure built for a different era, that AI builders are becoming infrastructure companies, and that some are tackling power and telecom problems themselves.
Why does optical networking matter for AI?
AI clusters move huge amounts of data between accelerators, memory and switches. Optical links, and increasingly co-packaged optics, carry that traffic more efficiently than copper at high speeds. We found no verified market-size figure, so we do not quote one.
Which AI models launched most recently?
Release trackers list Claude Haiku 5.5 on October 7, Mistral Large 4 in public preview on October 6, Gemini 4 Argon on September 30 and GPT-6.1 Sol on September 29, among others.
What does Jensen Huang say about photonics?
Huang told CNBC on May 29, 2026 that the silicon photonics capacity NVIDIA needs is substantially higher than the world has today. NVIDIA also says photonics switches can open the gate to million-GPU AI factories. Both are company statements.
Is the AI buildout a bubble?
Opinions differ. A news.com.au headline records Steve Eisman warning about the AI boom, while Meta’s October 22 capex raise shows at least one giant doubling down. Neither proves the outcome.
How much water do data centers use for cooling?
Rystad Energy estimates data centers used 222 billion liters worldwide for cooling in 2025, and that figure could reach 644 billion liters by 2030 without adaptive measures.
What should developers do about capacity risk?
Keep at least one fallback model and provider, track pricing changes, and avoid designs that assume a single vendor’s quota stays constant.
