Is AI Too Important to be a SaaS Purchase?
AI & Technology
As companies start token-maxing their Claude, ChatGPT and Copilot subscriptions, the old build-vs-buy conversation is back — is AI too important, and too expensive, to outsource?
For the last twenty years, companies have wrestled with one question about their technology stack over and over: buy (rent) or build? I have led Software-as-a-Service companies for more than 20 years — before SaaS was even talked about as a thing; I think we called it a "platform company" back in the day. That argument, with respect to software, was broadly won. Almost every company on the planet has stopped running its own email servers, data centres or CRM development programmes and just subscribed — Software-as-a-Service won.
Today, the issue is about Artificial Intelligence as a Service (AIaaS). I think the answer is not so clear cut, because intelligence — artificial or not — really should be a company's competitive advantage and strategic moat, and therefore should be home-grown, not outsourced.
Two stories this week brought this issue into sharp focus.
- Two powerful AI models are withdrawn from use. A US export-control order forced Anthropic to switch off its most powerful models, Fable 5 and Mythos 5. While both are still in a kind of market-testing/trial phase, imagine if they were already embedded into your entire technology infrastructure. Your business would shut down straight away.
- A major law firm announced plans to build rather than buy. Law giant Kirkland & Ellis is spending $500m over three to four years to develop its own AI platform, to ensure it maintains a clear, distinctive offering over other law firms who use standard AI services such as Harvey and Legora — which will inevitably spit out more homogenous advice for their client base, because it is the same models processing similar datasets.
On top of this, the cost of Artificial Intelligence is climbing — not because unit costs are rising (they are actually falling, ~80% in a year) but because total AI bills are ballooning as usage explodes. Average company spend leapt from roughly $63k to $85.5k a month in a single year, according to a CloudZero report last year — and no doubt this number continues to climb in 2026.
| Buy (ChatGPT / Claude / Copilot) | Build (open-source / your own) |
|---|
| Speed | Live in days | Months of engineering |
| Upfront cost | Low — pay as you go | High — talent + infrastructure |
| Control | Their rules, their outages | You hold the keys |
| Data privacy | Sent to a third party | Stays in-house |
| Upkeep | They handle it | You handle it (forever) |
| Best for | Most firms, speed wins | Where AI is your edge |
Why just buy?
For most companies, renting is the smart first move:
- Speed. You're live in days, not months — no GPUs to wrangle, no hiring spree.
- Low cost to start. Pay-as-you-go, no big upfront bet, easy to switch off.
- Someone else does the heavy lifting. Updates, security and scaling are their headache, not yours.
- Lower risk. If a model doesn't suit, you swap providers and walk away — no stranded kit.
Why build your own?
Building earns its keep only when AI is part of your edge, not just plumbing:
- Competitive advantage. A model shaped by your own data and know-how does things rivals simply can't copy — exactly Kirkland's bet, baking its lawyers' "collective intelligence" into a tool no competitor can licence.
- Control and privacy. Your data never leaves the building — often non-negotiable in healthcare, finance and law.
- No one can pull the plug. Own the stack and no government letter or provider outage takes you offline (see Fable 5).
The money question: when does building beat buying?
A rough rule of thumb — and it is only that. By one estimate from AI cost-analytics firm TokenMix, self-hosting only starts to make financial sense once your AI spend pushes past roughly $20,000 a month; below that, the engineers, servers and 3am call-outs usually cost more than you'd save. But the figure swings widely with which model you're replacing and how heavily you use it — so treat it as a signpost, not gospel, and I am not convinced this estimate is a fully loaded cost. The real crossover for your business turns on sustained volume running into the tens of millions of tokens a month.
And mind the difference between the two reasons to build. Self-hosting to save money is a maths problem with a clear break-even — but it needs to be fully costed. Building a platform to win your industry (the Kirkland move) is a strategic bet that can run to tens of millions, if not more — and it isn't about cheap tokens at all. It's about owning something rivals can't copy.
The honest answer? Most firms rent first, then build where it gives a genuine edge. You don't need a $500m war chest — just clarity on where AI is core to your business, and where it's merely plumbing.