The Compute Class Divide
Anthropic and OpenAI are in an arms race for enterprise customers. Last week, Anthropic announced organisation-wide controls for Claude Cowork. OpenAI slashed the price of its Pro subscription to undercut Claude Code. Both companies are building sales teams, compliance features, and dedicated capacity for the customers who can sign six-figure annual contracts.
This makes perfect sense for them. Enterprise is where the revenue is. Enterprise is how you justify $450 billion in global AI infrastructure spending.
But compute is finite.
The supply side nobody's talking about
Data centre GPU lead times are running 36 to 52 weeks. The hyperscalers (Microsoft, Google, Meta, Amazon) placed multi-billion-dollar forward orders for NVIDIA's Blackwell chips in 2025, consuming most available allocation through the end of 2026 and into 2027. Even NVIDIA's own research teams can't get enough GPUs internally.
The binding constraint isn't even the GPU die. It's the memory. HBM packages are crucial for AI accelerators, and the three companies that make them (SK Hynix, Samsung, Micron) are prioritising data centre orders. DRAM supplier inventories dropped from 13-17 weeks in late 2024 to 2-4 weeks by October 2025. Samsung raised certain server-memory prices by 30 to 60 per cent.
When a bank or insurer locks in dedicated AI capacity, that's capacity that isn't available to the rest of us. A hyperscaler will always allocate reserved compute to a $100M enterprise customer before making on-demand inventory available to a startup. That's not malice. That's arithmetic.
The two-tier model
Here's what this looks like in practice.
Enterprise gets Opus. The solopreneur gets Haiku. Enterprise gets dedicated capacity, custom rates, 99.99% uptime SLAs, and priority allocation. The rest of us get rate limits, usage caps, and whatever model sits in the budget tier.
The price gap is already visible. Enterprise Anthropic contracts average around $85,000 per year. A solopreneur on the Pro plan pays $240 a year. That's a 350x difference in what you're paying, and the allocation priority follows the money.
The optimistic version of this story is that models keep getting cheaper and more efficient. Opus 4.6 costs a third of what Opus 4.1 did. That's real progress. But the pessimistic version, and I think the more honest one, is that the frontier keeps moving. The best models get better, the compute to run them gets scarcer, and the gap between what enterprise can afford and what the rest of us can access doesn't close. It widens.
Democratisation was the pitch
The AI industry sold itself on democratisation. Anyone can build. The barriers are gone. The tools are free or nearly free.
Some of that is true. I'm a solopreneur. I've shipped ten products in six months using AI tools that didn't exist two years ago. I don't need a team of twenty developers. That part is real.
But "anyone can build" is not the same as "anyone can compete." Building is the easy part now. Competing requires the best models, sustained compute access, and the ability to run workloads at scale without your costs spiralling or your provider throttling you because an enterprise customer needs the capacity.
The Rubin Affordability Paradox, a term I saw floating around this week, names the demand side of this neatly. The firms that most need AI to stay competitive are least able to afford it. But there's a supply side to the paradox too: the infrastructure providers are actively optimising away from serving small operators. Not because they're hostile. Because the economics push in one direction.
What this means if you're small
I'm not going to pretend I have the answer to a structural supply constraint. I don't. But I've been thinking about what the solopreneur's actual advantages are when compute is king and you don't have any.
First: you don't burn compute on organisational overhead. No meetings about meetings. No compliance theatre. No 40-person Slack channel debating which model to use. Every token goes to output. A solopreneur at 95% utilisation beats an enterprise team at 40%.
Second: you can switch. When Opus is too expensive, you drop to Sonnet. When Anthropic raises prices, you route through OpenRouter. When a new open-source model matches last year's frontier, you run it locally. Enterprises can't pivot that fast. Their procurement cycles are measured in quarters. Yours is measured in minutes.
Third: you can be smarter about what you use compute for. Prompt caching, batch processing, model routing. These aren't enterprise-only techniques. A solopreneur who understands the cost structure of API calls can run a workload at a fraction of the naive price. The 90% savings from prompt caching plus the 50% from batch processing means a well-optimised small operation can get close to enterprise unit economics on specific tasks.
But I'd be lying if I said that was enough. If AGI arrives, or even if frontier models just keep widening the capability gap, and only organisations with $50,000+ annual contracts can afford full access, the rest of us are running the intellectual equivalent of a dial-up connection while enterprise gets fibre.
The honest question
I don't know how this plays out. Maybe open-source models close the gap fast enough that it doesn't matter. Maybe compute costs fall faster than demand rises. Maybe the whole thing equalises in three years and I'll look back at this post and laugh.
Or maybe we're watching the early formation of a permanent compute underclass. People who build with AI but never have access to the best of it. Always one tier behind. Always making do.
I'd rather see that clearly than pretend the playing field is level.
If you're a solopreneur or small operator navigating this, I'd genuinely like to hear what you're seeing. Are you feeling the squeeze yet, or is this still abstract?
Get in touch: jamiewatters.work · X/Twitter · LinkedIn