Microsoft buried the Copilot+ PC badge — 30 million paid seats tell another story

What Microsoft changed, and what it did not

The "Copilot+ PC" label was never a product. It was a hardware certification under Microsoft's Copilot+ PC specification: at launch, a laptop needed a neural processing unit capable of at least 40 TOPS — trillion operations per second — to qualify, enough to run a handful of features locally, such as Recall, Live Captions and Cocreator. Some newer reports cite 45 TOPS, apparently referring to newer device specifications or requirements, but that should not be confused with Microsoft's original Copilot+ PC threshold, which was 40 TOPS. Neither figure is a broader AI PC industry definition; other vendors use different numbers for similar marketing language. The NPU threshold had almost nothing to do with the Copilot assistant itself, which lives in the cloud. It was a line drawn to sell new silicon.

According to XDA Developers, Microsoft has removed that badge from new Surface devices, retired the mascot Mico and shown a revamped Copilot assistant. If you bought a 40 TOPS machine, you lost a sticker, not a feature set.

The 40 TOPS rule aged badly, and that is the point

Two years after launch, the NPU requirement looks like the least interesting number in the stack. Most Copilot work happens on Microsoft's servers: drafting, summarising, reasoning over your mailbox. The NPU handles a small set of local effects. It was never the reason enterprises signed contracts, and it was never the reason consumers got excited.

The more honest comparison is with hardware people actually buy to run models themselves. In AI Arena we test local inference on an NVIDIA RTX 5060 Ti 16 GB via Ollama, and the trade-offs are completely different from an NPU: 16 GB of VRAM puts a hard ceiling on model size and quantisation, power draw is measured in hundreds of watts rather than single digits, and you get far more raw compute in return. An NPU is a power-efficiency play. A GPU is a capability play. Microsoft sold the first as if it were the second.

30 million seats: do the arithmetic

The number that matters is not the badge. It is 30 million paid Microsoft 365 Copilot seats, a figure attributed to Microsoft AI executive Nicolas Bustamante's comments on X and cited in XDA Developers' report. To be clear, the dollar figure that follows is a hypothetical gross list-price ceiling, not reported Microsoft revenue. The $30 per user per month list price applies to the relevant Microsoft 365 Copilot licence, not to every Copilot-flavoured seat; the 30 million figure can include different licensing arrangements, discounts and bundles. If every one of those seats paid the full $30 list price, the total would be roughly $900 million per month, or about $10.8 billion annualised — the ceiling, not the income line. In euro terms, using a rounded illustrative exchange rate rather than a quoted Microsoft price, that is roughly €27–28 per seat per month before VAT — so a 250-seat deployment at list is a commitment in the range of €80,000 a year before any enterprise discount. Procurement teams rarely pay list, but the order of magnitude is what budgets are built on.

It is worth sanity-checking what $30 buys elsewhere. At currently published API prices for large commercial models, the same money can buy tens of millions of input tokens or millions of output tokens, depending on the model, tier and whether the customer qualifies for discounts. Specific list prices move quickly, so the useful comparison is not a single model rate; it is that raw token capacity is now an order of magnitude cheaper than a per-seat assistant. The caveat is obvious and important: a Copilot seat is not a token bundle. You are paying for identity integration, tenant data controls, audit logs and support — things no raw API gives you. But it frames the question every buyer should ask: of those 30 million seats, how many are opened weekly?

Why enterprises actually stay

Nicolas Bustamante's argument, made publicly in a post on X and reported by XDA Developers, is not that Copilot is the smartest assistant on the market. It is that large companies rely on it for data privacy and compliance. He also wrote that he went from knowing "maybe five people" using Copilot to seeing it widely deployed across enterprises — a shift that is plausible and easy to explain.

The alternative to a governed assistant inside your Microsoft tenant is not nothing. It is employees pasting contracts, salary tables and internal roadmaps into consumer chatbots. From a risk officer's perspective, a slightly worse assistant with DLP policies and an audit trail beats a better one with no controls at all. That is the real moat, and it has nothing to do with NPUs or mascots.

The European angle: enforcement is no longer theoretical

EU AI Act deadlines have hit in waves, not all on one date. The Article 4 AI-literacy obligation has applied since 2 February 2025. The general-purpose AI model obligations have applied since 2 August 2025. The Article 50 transparency obligations apply from 2 August 2026 and target specified AI-generated or manipulated content — for example, synthetic audio, video or images covered by the regulation, and certain AI-generated text where the regulation requires disclosure. Whether a company's Microsoft 365 Copilot deployment triggers them depends on the company's role and the use case, so they are not automatic for every Copilot seat. For European companies running Copilot, that means taking a case-by-case view: internal policy on disclosing AI-generated material in customer-facing output and documented staff training. Maintaining a register of which AI systems are in use and for what is a recommended governance step, not a universal Article 50 requirement. The legal obligations and sensible internal controls should not be conflated.

On the data side, Microsoft states that its EU Data Boundary covers Microsoft 365 Copilot processing for European customers. That is Microsoft's compliance argument — not a standalone regulatory guarantee — but it is the reason DPOs in Frankfurt, Paris and Prague keep asking about it. The European Commission's regulatory framework is the reference point for what your DPO will ask about.

European alternatives, honestly assessed

Mistral's EU-hosted commercial models remain a credible option, with competitive API pricing and a data-residency story that is simpler to explain to a regulator. What they do not have is deep integration into Outlook, Teams and SharePoint — and that integration, not model quality, is what people are actually paying $30 a month for. Anyone claiming a European switch is painless is selling something.

If you are paying for Copilot seats

Three things are worth doing this quarter. First, pull your usage report and count monthly active seats against paid seats — Copilot's enterprise economics live or die on that ratio. Second, test whether the workloads you actually bought it for — meeting summaries, drafting, document Q&A — are still worth the seat price now that API pricing has fallen to cents per million tokens; a small internal tool with a governed API key may cover 80 percent of the demand. Third, check whether Article 50 obligations apply to your use cases before your next audit, not after it.

And if you are a power user rather than a procurement officer, the more interesting experiment is still local. A capable open-weight model on a 16 GB GPU will not match Copilot's Office integration, but it will answer a lot of questions without sending your data anywhere. That is the measurement we keep running — and one of the reasons we test in our own rig rather than repeating press releases.