Twenty-five years ago we designed a network to consume expensive external resources deliberately. AI now requires the same discipline.
During the dot-com era, I was the first employee of , Perth's first broadband ISP (well, that's what we told everyone anyway).
Where it all ended up
WiredCity's internet division was absorbed by Amcom, the Perth fibre carrier, and folded into its Amnet ISP brand alongside Perth Internet and Arachnet. Amcom went on to buy just under a fifth of the competitor across the street.
- WiredCity ISP to Amcomc. 2004
- Amcom takes 19.97% of iiNet2006
- Amcom to Vocus, $687m2015
- iiNet to TPG, $1.56bn2015
- Vocus taken private2021
The bozos across the street outlasted us by a decade and were then eaten in turn. Every name on that list now sits inside something larger, which is the part worth keeping.
Amnet's acquisition history →This was the tail end of dial-up, smack in the middle of the dot-com bust, which is a polite way of saying the money had stopped. Those across the street bought bandwidth at a scale we could not match and priced accordingly. Our meagre 10 megabit upfeed came from Telstra, the monopoly of the day, which : it sold us the carriage we could not operate without, while competing for the very same customers. Every megabit cost us money we did not have.
They were not bozos
Michael Malone started iiNet in 1993 in the garage of his parents' house in Padbury, having worked out that finishing his degree at UWA would mean losing his internet access. There were no ISPs in Perth to go to, so he built one.
- Founded1993, in a garage
- Customers1.3 million
- Sold to TPG$1.56bn, 2015
It became the second largest ISP in the country, with more than two thousand staff, and an ASX listing to go with them. The bozos did rather well, and deserved to.
iiNet through the years →Supplier and competitor at once
Telstra owned the network, sold wholesale carriage to the ISPs it competed with, and then priced against those same ISPs at retail. That is a structural conflict rather than a character flaw. One company set its rivals' input costs and their market prices at the same time.
Australia spent the next twenty years trying to unwind it, through access regulation, then accounting and operational separation, and finally the structural separation that produced the NBN. It took a national infrastructure project to fix a supplier relationship nobody had governed at the outset.
The odds were not exactly stacked in our favour, so we did three things to control the economics.
Firstly, we installed a proxy server. If ten customers requested the same popular web page, the proxy retrieved it once, cached it and served the remaining requests locally. Instead of purchasing the same external bandwidth ten times, we purchased it once.
Secondly, we joined the , or WA-IX. When one of our customers requested content hosted by another Australian ISP, it made little sense to send that traffic through a paid upstream carrier, so we all connected to each other via peering agreements. The traffic stayed local, the route was shorter, performance improved and we saved a few dollars.
What an internet exchange does
An exchange is a shared piece of switching fabric where networks in the same city plug into each other and agree to swap traffic directly, instead of paying a carrier to haul it between them.
Without one, a request from a customer of one Perth ISP to a server hosted by another could travel to the east coast and back to cross a few suburbs. The commercial point is simple: traffic that stays inside the exchange never appears on an upstream bill.
Thirdly, we rate-limited the hell out of the software and music pirates. Peer-to-peer applications such as BitTorrent could consume a disproportionate amount of shared bandwidth while degrading the service for everyone else. Using Cisco's , or NBAR, that traffic was identified and slowed to a crawl. While it annoyed the hell out of said pirates, they were a minority and eventually moved elsewhere.
How NBAR saw what was on the wire
NBAR inspected traffic and identified which application had generated it, rather than trusting port numbers. Peer-to-peer software had learned to disguise itself by moving off its default ports, so ports had stopped being a reliable signal.
Once the router could name the traffic, policy could act on it: mark it, cap it, or push it behind everything else. The technology was classification. The purpose was deciding whose work got the expensive capacity.
These were technical decisions, but their purpose was commercial governance. We did not simply accept whatever traffic users and applications generated and then purchase enough upstream capacity to accommodate it.
We designed the network to consume expensive external resources deliberately.
AI now requires the same discipline, with the new unit of account: .
What you are actually buying
A token is a fragment of text, roughly a short word or part of one. Models read and write in tokens, and providers bill by them, counting both what you send and what comes back.
That makes the meter a measure of model activity rather than of anything the business values. A long prompt that produces a wrong answer costs more than a short one that produces a right answer. Nothing in the pricing knows the difference.
Never Let a Supplier Control Both Price and Consumption
Tokens, just like before them, are a measure of activity, not a unit of business value. The latter is a very different thing, and good governance means avoiding situations in which a supplier effectively controls both the unit price of an activity and how much of that activity you consume.
What happened to the price of bandwidth
Wholesale IP transit, the thing an ISP buys by the megabit, has been tracked since 1998. The direction has never reversed.
- 1998$1,200 per Mbps
- 2010$5 per Mbps
- 2022$0.09 per Mbps
A fall of more than thirteen thousand times in twenty-four years, while the volume carried grew every year of it. Nobody counts bandwidth any more. The price of it fell so far that the meter eventually stopped being worth reading.
The first two figures are William Norton's, gathered from operators at peering conferences. The last is TeleGeography's, for the cheapest 10 GigE ports in the major hubs.
TeleGeography on transit price erosion →I always assumed that in the enthusiasm for AI we would do things inefficiently in the early innings. We saw increasingly large prompts, growing context windows and agents that make repeated model calls, so a relatively simple business process can generate an extraordinary number of tokens before producing a useful outcome. In a learning phase that is the price of finding out. Once AI becomes operating infrastructure it is weak governance, as have discovered.
Discovering it the expensive way
Uber burned through its entire 2026 AI coding budget in four months, after an internal leaderboard encouraged engineers to take the tools up. About a tenth of its committed code is now written by autonomous agents.
- Uber's 2026 AI coding budgetgone by April
- One firm, usage uncapped$500m in a month
Amazon's chief technology officer, Werner Vogels, named the pattern at the UN's AI for Good summit: "We see a shift happening between the cheaper open source models and the bigger expensive models." Asked whether you need the highest-end model to solve a given problem, his answer was that you do not.
None of these are stories about AI failing to work. They are stories about consumption nobody was governing.
Fortune, July 2026 →So, are there any twenty-five year old dot-com era lessons that still apply here? I think there are three.
- Proxy. Only send what you need to.
- Peering. Be judicious with the remainder.
- Policy. Ask whether you need a model at all.
Each is a way of designing a solution that stays resilient to whatever its supplier does next, and all three are practical design choices you can make today.
The Proxy: Stop Buying the Same Answer Twice
Start here, because it is the easiest of the three to apply. The principle transfers directly to AI-generated responses.
Many applications repeatedly ask a model questions whose answers change rarely or not at all. Corporate strategies and priorities tend to follow planning cycles, and a policy explanation may remain valid for months. Lessons learnt from prior projects may already be captured. A document may have been summarised previously.
There is no reason to regenerate these answers on every request. The system can:
- cache approved responses;
- store previous classifications;
- reuse extracted document fields;
- maintain embeddings locally;
- precompute common outputs;
- retrieve authoritative content;
- invalidate cached results when the source changes.
This is not merely a cost optimisation. It improves consistency. If an employee asks the same policy question twice, producing two subtly different answers is often a defect rather than a feature.
Peering: Stop Sending Everything Upstream
Peering was about finding a cheaper route for traffic that never needed the expensive one. In AI that means model routing, local execution and open weight models: sending each piece of work down the cheapest path that can actually carry it.
This is an area of extraordinary competition, and it is worth noticing who has suddenly taken an interest in the orchestration layer. NVIDIA has , the repository the open model world runs on, for close to $13 billion. Microsoft's exclusive arrangement with OpenAI ended in April, it is building its own , has joined NVIDIA and Meta in against restrictions on open weight models, and even publishes Phi, its family of , under an MIT licence.
Microsoft's own models
MAI is Microsoft AI's in-house model family, run by Mustafa Suleyman, who co-founded DeepMind and then Inflection. The first two, MAI-Voice-1 and MAI-1-preview, arrived in August 2025.
By April 2026 the family covered speech, transcription and image generation, with MAI-Voice-1, MAI-Image-2 and MAI-Transcribe-1 released for commercial use through Microsoft Foundry.
The point is not that any of them beats a frontier model. It is that a company which spent years reselling somebody else's now has its own, and only a non-exclusive licence to the other.
Microsoft on the MAI models →Reported, not yet confirmed
The Information reported in late August 2026 that NVIDIA had agreed to acquire Hugging Face for $12.9 billion, and CNBC, Fortune and others followed. Neither company has confirmed it, and Reuters reported that no agreement had been signed and that the talks could still collapse.
We cite it anyway because the direction of travel is the point. A company whose margins come from frontier training does not go courting the open model repository unless it expects open models to matter.
What they actually asked for
Two dozen companies signed an open letter on 24 July 2026, among them Meta, Microsoft, NVIDIA, IBM, Dell, Palantir, Hugging Face, Mistral, the Linux Foundation and Mozilla. They asked Washington to expand compute access for startups and researchers, fund shared training datasets and evaluation frameworks, and avoid what they called premature restrictions on open models.
It landed a week after Moonshot AI released Kimi K3, an open weight model said to rival the best from OpenAI and Anthropic, which rattled investors and had policymakers talking about restricting Chinese AI. "Our AI leadership will be judged not by one frontier AI model," the letter argued, "but by whether the United States builds a strong, open ecosystem."
The case they made for open weights was that these increase competition and prevent vendor lock-in. Which is this article's argument, made by the vendors.
Fortune on the open-weight letter →The cheaper route, sold by the incumbent
A small language model is built to do bounded work reliably on modest hardware, rather than to compete at the frontier. That is the whole proposition: not the most capable model available, but one sufficient for the task, running somewhere you control and pay for by the hour.
Phi is Microsoft's family of them, released under the MIT licence and available through Azure AI Foundry, Hugging Face and Ollama. Foundry Managed Compute adds a curated catalogue of open weight models from the Hugging Face ecosystem, deployable onto compute billed by the hour rather than by the token.
The catalogue is not limited to Microsoft's own work. It carries models from DeepSeek, Qwen, Moonshot, MiniMax and OpenAI's gpt-oss line, which is a striking list for a company whose frontier position depends on the opposite arrangement.
Hugging Face models on Foundry →Amazon and Google are no different. Bedrock in a single release in February, among them DeepSeek, Kimi, Qwen and GLM, which AWS itself describes as "delivering frontier-class performance at significantly lower inference costs". Google released in April under an Apache 2.0 licence, the first of that family to drop Google's own restrictive terms, built on the same research as Gemini but packaged to be downloaded and run wherever you like.
Six in one announcement
On 10 February 2026 AWS added DeepSeek V3.2, MiniMax M2.1, GLM 4.7, GLM 4.7 Flash, Kimi K2.5 and Qwen3 Coder Next to Bedrock as fully managed models, describing them as bringing "access to the most capable open weights models available today, delivering frontier-class performance at significantly lower inference costs".
All six come from Chinese labs. The cheapest sufficient path increasingly runs through models the American frontier labs did not build, which is some of the context for that open letter.
AWS on the six models →The licence is the story
Google DeepMind released Gemma 4 on 2 April 2026: four open weight models built from the same research as Gemini 3, running from small edge variants up to a 31 billion parameter dense model.
- LicenceApache 2.0
- Sizes2B to 31B
- Context256K tokens
Earlier Gemma releases carried Google's own terms, with monthly active user limits and an acceptable use policy attached. Gemma 4 is the first of the family under OSI-approved Apache 2.0, which means no user caps and no strings on commercial use.
A company does not relax its own licence unless it wants the models to spread.
Google on Gemma 4 →NVIDIA is making the same bet in hardware. Its is basically a desktop supercomputer for local models of up to 200 billion parameters. Why sell these when the hyperscalers are buying your chips with your on top? Well, if routine intelligence goes local and commoditised, like the Gemma 4 example, NVIDIA intends to sell the machines it lands on.
Seventy-five per cent
- Q4 FY2026 gross margin75.0%
- Full year FY202671.1%
- Q4 data centre revenue$62.3bn
Those are NVIDIA's own GAAP figures for the quarter ended 25 January 2026. The full year sits lower because the first quarter did.
A margin like that is an invitation. It funds the customers designing their own silicon to get out from under it, and it funds the competitors. Which is the reason to own a position in whatever comes next, including the cheap end.
NVIDIA Q4 FY2026 results →A frontier-ish model, on a desk
- Unified memory128 GB
- Inference ceiling200bn parameters
- Form factorDesktop
The number that matters is the memory, not the raw throughput. Unified memory is what determines whether a model fits on the machine at all, and a model that does not fit cannot be run locally at any speed.
NVIDIA DGX Spark →They are all hedging, in other words. So should you.
Policy: Start With Determinism
Proxy is all about only sending what you need to, and peering is then about being judicious with that remainder. Policy is where NBAR comes back: it named what was actually on the wire so we could decide what deserved the expensive link. Apply this to AI and the question is whether a model is needed at all.
Language models are probabilistic. They generate likely responses rather than execute a guaranteed sequence of instructions. That is valuable when a problem contains genuine ambiguity that cannot be reduced to a series of "If, then" statements. It is unnecessary, and sometimes dangerous, when the correct behaviour is already known.
Consider invoice processing. During the initial wave of AI adoption, I saw a lot of very elaborate prompts such as the one below. Not only is it brittle and time consuming shoe-horning the agent into acting deterministically, it is token heavy. Yet invoice processing is a mature business process, and when you look at the steps involved, much of the work can be performed locally by workflow tooling that most organisations are .
The engine you already licensed
Power Automate is seeded into Microsoft 365 Business and Enterprise plans, covering flows across the standard connectors: Outlook, SharePoint, Teams, OneDrive, Excel and Forms. ServiceNow's Flow Designer ships with the base Now Platform and is included in the platform licence.
Both have a boundary worth knowing before anyone calls them free. Power Automate's premium connectors, SQL Server, Salesforce, HTTP and anything custom, need a paid licence. ServiceNow charges separately for Integration Hub once a flow reaches outside the platform.
Even so, an invoice approval that never leaves the tenant is running on something already bought and paid for, and it will not send you a bill per token for doing it.
Power Automate licensing FAQ →While AI can seriously improve the process via reading an unfamiliar document or interpreting an unclear description, I'm not convinced of its utility to decide whether the accounting entry balances or whether a known approval threshold has been exceeded. A frontier model may perform all these tasks impressively. That does not make it the correct tool.
Using frontier intelligence to apply a known business rule is like hiring Taylor Swift to record your on-hold music. I have no doubt people waiting in the queue for a human will enjoy the bangers, but the decision remains economically absurd.
Good design asks: where does genuine uncertainty enter the process? Use AI there. Everything before and after that point should remain as deterministic, testable and auditable as practical.
Three Lessons, One Principle
The emerging AI architecture combines all three lessons from WiredCity.
Distilled to a principle, we could call it this.
Design for the least sophisticated system capable of producing a reliable result.
Or, in the language of our ISP strategy: reuse what is already known, route work through the cheapest sufficient path, and prevent low-value activity from consuming premium capacity.
Who Captures the Value Is Up to Us
Frontier-model providers may continue to produce the world's most capable models. Hyperscalers will remain essential for training, elastic capacity, variable workloads and large-scale deployment.
But neither fact guarantees unlimited pricing power. If they end up with it anyway, that is on us.
Once AI consumption becomes a material operating expense, we will optimise it. We will shorten prompts, cache stable answers, separate rules from reasoning, route work between models and process suitable tasks locally. If we instead embed our entire operating process into a proprietary model's ecosystem, substitution becomes harder.
Proxy servers did not eliminate the internet. Peering did not eliminate upstream carriers. NBAR did not eliminate file sharing. They allowed us to reserve expensive shared capacity for the traffic that justified it.
The lesson for AI is becoming clear.
Do not use sophisticated intelligence where a known answer, deterministic workflow or commodity model will do.
That is not resistance to AI. It is competent governance.
This piece extends arguments we made earlier in AI Works. The Returns Might Not. and We Built a CRM. The Bond Market Will Finish the Job.
