Local AI or Seat Subscriptions: The Pricing Just Moved on Both Sides
Synergised Consulting Ltd
Last Updated: 26 October 2026
Two pricing moves landed this month that change the buy-versus-subscribe maths for AI work in a small business, and they moved in opposite directions. OpenAI added Premium seats to ChatGPT Business at $100 per user per month on annual billing [1], while NVIDIA raised the 128GB DGX Spark, its desk-side AI computer, to $6,950 [3], roughly 75 per cent above where that machine sat a year ago [4]. Whichever route you were quietly leaning towards, the ground under it shifted. This piece sets out what each side now costs, what a realistic mid-size team would actually pay, and the one document worth building before you commit either way.
What Changed, and When
Both vendors repriced within weeks of each other, which is why the decision feels different now than it did in the summer. On the subscription side, OpenAI introduced Premium seats for ChatGPT Business, available from 25 August 2026, at $125 per user per month billed monthly or $100 billed annually, carrying five times the usage of a Standard seat with no five-hour usage cap [1]. Standard seats stayed put at $25 monthly or $20 annual [1]. Anthropic's Claude Team plan mirrors that structure almost exactly: Standard seats at $25 monthly or $20 annual, Premium at $125 or $100, with a minimum of two seats and a ceiling of 150 [2].
On the hardware side, the 128GB DGX Spark launched in October 2025 at $3,999, rose to $4,699 in February 2026, and now lists at $6,950 on NVIDIA's own store [5]. A cheaper 64GB model arrives on 23 October at $4,999, sold through Acer, Asus, Dell, Gigabyte, HP and MSI rather than by NVIDIA itself, and there is no NVIDIA-branded version of it [5]. Memory prices are the stated cause, driven by data centre demand [4].
The capsule version: subscriptions gained a pricier, heavier tier for power users, and the flagship local box became 75 per cent more expensive in a year. Neither side got cheaper for the workloads that matter most.
What Seats Actually Cost a Real Team
Seat pricing looks small per user and grows linearly with headcount, which is the property worth sitting with for a moment. At OpenAI's published rates, a Standard seat is $20 per user per month on annual billing and a Premium seat is $100 [1]; Anthropic's Claude Team matches those numbers seat for seat [2]. For a 25-person firm where everyone gets a Standard seat, that is 25 seats at $20 for 12 months, or $6,000 a year per vendor. Run both vendors side by side, as many firms quietly do, and the baseline doubles before any API usage beyond the included limits, which both vendors meter separately [1][2].
The new Premium tier changes the shape of that bill rather than just its size. A realistic mixed allocation for a 25-person firm might be 22 Standard seats and 3 Premium seats for the two or three people doing the heaviest work. On OpenAI's annual rates that is (22 x $20 + 3 x $100) x 12, which comes to $8,880 a year from one vendor [1]. The identical seat mix on Claude Team prices out the same way, $8,880 [2]. The lesson is not that this is unaffordable; it is that the bill is a function of headcount and seat mix, so it reappears every month and moves whenever either vendor reprices.
The Dependency Nobody Counts
There is also a dependency hiding in the seat model that has nothing to do with price. Every workflow that runs through a vendor's seats stops the day the seats lapse, whether that is a billing dispute, a card expiring or a product change. Counting how many of your actual workflows sit behind each vendor's login is part of the register described below, and in most firms that count has never been written down anywhere.
What the Spark Price Rise Changes
The local option was the hedge against exactly this subscription creep, and the hedge itself just got more expensive. The 128GB DGX Spark listed at $3,999 when it launched in October 2025, moved to $4,699 in February 2026, and now lists at $6,950, which The Register puts at nearly 75 per cent above where it stood a year earlier, attributing the rise to memory supply being pulled by data centre buildouts [4][5]. NVIDIA has also introduced a 64GB version at $4,999 through partner manufacturers, shipping from 23 October, with half the memory of the flagship [4]. Two clustered 64GB units recover the memory of one 128GB machine but cost $9,998 to do it [5].
Against the seat bill, the arithmetic still reads a particular way. One 128GB Spark at $6,950 is a one-off cost, while a 25-seat Standard deployment on one vendor runs about $6,000 a year [1], so the hardware reaches price parity with roughly fourteen months of seating and then keeps running. That comparison is deliberately simple, and the next section shows why it is also incomplete: the two options buy different things, and neither replaces the other for most of what a firm actually does.
What Each Route Actually Buys
A subscription seat buys breadth: every person in the firm gets general-purpose AI help with no infrastructure to own. A local box buys depth and control: it runs specific workloads on hardware you own, with the data never leaving the building, but it does nothing for the twenty-two people who just want better drafts of things. Treating them as substitutes is the mistake; they are different purchases for different parts of the workload.
To make that concrete, take one illustrative scenario built entirely from the published prices above. A 25-person accountancy practice pays $6,000 a year for one vendor's Standard seats [1], and its tax team of three upgrades to Premium seats, lifting the bill to $8,880 [1]. The same firm also holds client documents it would rather not upload to any cloud service, so it buys one 128GB Spark at $6,950 [3] and runs document processing locally on it. Total first-year cost is about $15,830, and the firm now owns a machine that does not expire. Strip the local box out and the firm saves the capital but sends every sensitive document through a vendor it does not control; strip the seats out and the twenty-two other staff lose their daily tool. Neither purchase covers the other's job, which is exactly why the comparison cannot be settled by price alone.
The general shape, stated plainly: subscriptions are the right default for broad, low-stakes, everyday use across a whole team, and local hardware earns its place when a specific workload involves data that should not leave the premises or usage volumes that would blow through seat rate limits [1]. Most firms end up needing some of each, which is why the decision worth documenting is allocation, not allegiance.
The Question to Ask Before Either Purchase
In discovery workshops, the list of what a firm actually uses AI for gets built in conversation rather than handed over: walking a normal week task by task, the everyday drafting and summarising separates itself quickly from the handful of workflows that touch client data or run in volume. Those two groups almost always point to different answers on the buy-versus-subscribe question, and owners rarely arrive knowing which of their workflows fall in which group. That mapping is the part nobody sells you, and it is the part that decides the purchase.
Document It Before You Decide
The decision that fits this moment is not "subscribe" or "buy hardware", it is document. Build one register before the next renewal date lands: for each AI tool in the business, list the seats, the monthly or annual cost, which workflows depend on it, and when the renewal falls. It is a morning of spreadsheet work, and it converts a vague sense that "AI costs are creeping up" into a number you can act on, especially in a quarter when both sides of the market repriced.
The register earns a second use beyond cost control. It shows exactly which operational workflows rest on which vendor's seats and which, if any, run on hardware you own, and that is the kind of record a buyer's diligence process can inspect directly: it answers, with evidence rather than assurances, what actually happens here when a vendor changes its prices or terms. In a market where both sides have just repriced, the business that holds that record is the one that can react in a week instead of a quarter.
Sources
- [1] OpenAI Help Center, ChatGPT Business overview:
- [2] Anthropic Help Center, What is the Team plan:
- [3] NVIDIA Marketplace, DGX Spark product page:
- [4] The Register, Nvidia debuts $4,999 DGX Spark with half the RAM and storage, 2 October 2026:
- [5] Implicator.ai, Nvidia adds 64GB DGX Spark at $4,999 as 128GB model climbs to $6,950, 4 October 2026:
