What we will and won’t claim here
AI is full of big claims about speed, price and efficiency, or about being first of its kind. Most of them come without a date, a source or a way to check.
We'd like our News and Research sections to work differently. This page sets out the rules we follow, so you know what to expect from us and what you can hold us to.
Every number has a source
If we publish a number, we tell you where it came from.
That means one of three things. Either we measured it ourselves and say how. Or we read it from a public source and link to it. Or it comes from someone we interviewed, and we say who said it.
If we can't point to one of those, the number doesn't go in. That rule costs us some good-sounding sentences. We think it's worth it.
Every number has a date
Prices, speeds and availability in AI change quickly. A figure that was true in spring can be wrong by autumn.
So every number we publish carries the date we read or measured it. Our Model Watch posts, for example, list prices with the day they were read. If you come across one of our pages a year from now, you will know how old the information is and can judge it for yourself. We explain the thinking in why we publish dated prices.
When newer figures replace older ones, we say so on the page and keep the earlier figure, with its date, in the corrections log or the previous post. You should always be able to see what changed.
Other people's numbers stay theirs
Much of what we publish is about the wider market: what models cost, how fast they run, how capacity is being built. Those are other companies' figures.
When we report them, we say so. We don't present someone else's list price as our price, and we don't turn a comparison into a claim that we are better. If a page is about the market, it stays about the market.
We say what we are building, carefully
Starfield is building products that are not live yet. When we write about them, we use words like "designed to" and "we're building." We don't publish performance figures for our own systems until they are running real workloads and we can measure them properly. We don't publish launch dates we can't stand behind.
When our own numbers arrive, they will follow the same rules as everyone else's: method, date and source.
What we leave out
Some things you won't read here:
- Superlatives and "first of its kind" claims we can't prove
- Customer names, unless the customer has agreed in writing
- Partners or deals that haven't been announced by both sides
- Funding, valuation or revenue figures
- Claims about our technology that our CTO hasn't checked
That last point is a working rule inside the company. Before anything technical goes on the site, Ian Arden, our CTO, checks it. If a claim isn't on the list of things we can stand behind, it doesn't run.
What we say, what we leave out
What you will read here
- Numbers
- Every number has a source you can check and a date.
- Our products
- Described with "designed to" and "we're building".
- Other companies
- Their figures, labelled as theirs.
What you won't read here
- Numbers
- Funding, valuation or revenue figures.
- Our products
- Performance figures before we can measure them properly.
- Other companies
- Customer names unless agreed in writing, unannounced deals.
Quotes are checked against the source
When we quote someone from an interview, podcast or panel, we check the words against the recording. If we tighten a quote for length, we say it has been edited. We give the timestamp where we can, so you can listen for yourself.
When we get something wrong
We will get things wrong. Prices change between reading and publishing. A figure gets copied badly. A source updates its page.
When that happens, we fix it in the open. We keep a corrections log on the Research section. Each entry says what was wrong, what we changed and when. We don't quietly edit a page and hope nobody noticed.
If you spot a mistake, tell us at hello@starfield.net, as the corrections log says. Please include the page, what you think is wrong and, if you have one, a source. We read what you send, and we update the log whenever we make a change.
What you can hold us to
Here is the short version, in one list:
- Every number has a source you can check.
- Every number has a date.
- Other companies' figures are labelled as theirs.
- Our products are described as what they are today, not what we hope they'll be.
- Quotes are checked against the recording.
- Mistakes are corrected publicly, with a date.
If any page on our site breaks one of these rules, that's a mistake, and we want to hear about it.
Why we bother
It would be easier to write the way much of the industry writes. Big numbers, no dates, bold promises. It would also make our pages less useful to the people we most want to reach: engineers comparing options, buyers writing business cases, analysts checking claims.
Those readers need numbers they can rely on. A dated, sourced figure can go into a spreadsheet or a board paper. A vague claim can't.
We also think trust is built the same way in publishing as in infrastructure. You show your work, keep records and make it easy for someone else to check. That is the mindset behind AI infrastructure operators, not developers.
What we're building
That same idea runs through our product. Starfield is building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. It is designed so every workload can generate an auditable record of commitment, placement and delivery, so buyers can check what they received.
Read our latest numbers on the Research page, or use Request capacity below to talk to us.
Related
Model Watch · AI infrastructure operators, not developers · LLM API pricing: why we publish dated prices