Everyone assumes the enterprise moved to the cloud years ago. Camberley Bates has spent three decades being paid to check assumptions like that, and this one is wrong. The Futurum Group VP and 30-year enterprise technology analyst joined Jeff Pedowitz on Unscripted to separate AI infrastructure reality from vendor hype. Her numbers will surprise anyone who thinks the on-prem era is over.
Who is Camberley Bates?
Bates is VP and Practice Lead at The Futurum Group and formerly led Evaluator Group, the Boulder-based infrastructure analyst firm, as Managing Director. She built and sold companies before becoming an analyst, and her entire job for 30 years has been cutting through vendor claims to tell executives what's actually real. She's also, by her own proud admission, a mainframe person.
What IBM's earnings dip actually means
When IBM's numbers disappointed the market, headlines read it as weakness. Bates read it as the opposite. Enterprises are pouring so much money into standing up AI infrastructure that they delayed mainframe upgrades, and mainframe revenue carries enormous margin. The money didn't disappear. It moved to the AI buildout, and the upgrade cycle will come back.
The context most people miss: mainframes still process most financial transactions and a huge share of airline bookings. The machines everyone forgot about still run the economy.
The stat that breaks the cloud-first narrative
Roughly 70 to 80 percent of the enterprise data center is still on-prem. Not legacy scraps. The core.
Cloud-first, the mantra of a decade ago, is dead. What replaced it is workload placement: for each application, where is the best home based on cost, governance, and performance? Stable applications often cost dramatically less on-prem. Repatriation is rare because it's brutally hard, but the default has flipped. Nobody assumes cloud anymore.
You can't outrun physics
Ocean data centers? Microsoft tried and quietly walked away. Data centers on the moon? Bates has a two-word answer: speed of light.
Compute has to sit near the data, and the response has to travel back. That's why the Gulf states' data center buildouts make sense (roughly 3 billion people within workable latency) and why US transactional AI keeps pulling workloads on-prem. It's also why the entire chip industry is obsessed with memory hierarchy: moving data ever closer to the GPU to keep it running at 100 percent. Data gravity, not ideology, is deciding where AI lives.
The prompt privacy problem nobody talks about
Here's the part of the conversation that should stop every executive cold. Bates was an expert witness in a court case and was barred from using AI, not because of the data, but because of the prompts. Your questions reveal your strategy. String someone's prompts together and you can reconstruct where they're going, what they're worried about, and what the numbers look like.
Jeff asked directly: if you're on an enterprise license that promises your data stays out of training models, is it guaranteed? Her answer: "I wouldn't." Thirty years of professional skepticism, applied.
This is what's driving the Dell and HPE boom. Enterprises are buying integrated on-prem AI stacks, training against models like Anthropic's on their own infrastructure, and running inference next to their transactional systems. The world is splitting into companies that own their AI and companies that rent it.
How to decode "AI-ready" and other vendor claims
After 30 years of hype cycles, Bates says this one isn't different in kind, just in speed. Her field guide:
"AI-ready" means nothing until scoped. Is it infrastructure integration? A data pipeline? The current tell is model flexibility: real AI readiness means plugging into whatever foundation model you need, with no lock-in.
Watch the roadmap language. When a vendor says a capability is 12 to 18 months out, ask whether that means beta or GA. The answer tells you whether any code exists yet.
Vendor websites got worse on purpose. A decade ago you could download the admin manuals. Now the detail is gated because marketing wants a conversation, not an informed download. And the manuals AI scrapes are frequently wrong, which is why competitive analysis still needs human interpretation.
On SaaS valuations and the seat-based model
Salesforce and its peers are posting healthy numbers while their stocks get hammered. Bates's read: the market isn't pricing the next 18 months, it's pricing the next 10 years, and it doubts the growth engine (more seats, more add-ons) survives a world where AI shrinks headcount. Her contrarian call: most of these companies adapt and come back. Not all. Like every disruption, there will be fallout.
The claim that will age worst
Jeff asked which AI prediction she's most confident will look foolish in a decade. No hesitation: the claim that AI eliminates work and we all live on universal income.
"People need work to have value in their souls and spirits," she said. Her proof point arrived the same week: her own public rant about ChatGPT's customer service. The AI company, of all companies, proving that people still need people.
FAQ
Is the cloud era over?
No. But cloud-first as a default is. Enterprises now place each workload where cost, governance, and latency make sense, and 70-80 percent of enterprise data center capacity remains on-prem.
Why are enterprises buying on-prem AI stacks?
Data control and prompt privacy. Training and inference next to transactional systems keeps sensitive data (and revealing prompts) out of public-facing models, and vendors like Dell and HPE are posting huge numbers because of it.
What does "AI-ready" actually mean?
Whatever the vendor wants it to mean. Scope it: infrastructure, data pipeline, or model flexibility. The strongest current definition is the ability to plug into any foundation model without lock-in.
Will AI wipe out SaaS companies?
Bates thinks most adapt and return over a 10-year horizon, though seat-based pricing is genuinely threatened and some players won't survive the transition.
Watch the full conversation on Unscripted with Jeff Pedowitz.
The Pedowitz Group | pedowitzgroup.com