Let us travel back. November 2024. AI is exciting and new and the bills haven't really arrived yet. The enterprises are signing contracts, standing up Copilot seats, wiring Sidekick into storefronts, and telling their boards a very confident story about competitive advantage. The tokens are flowing. Nobody is watching the meter.

Meanwhile, in a colocation facility somewhere, a server is humming. Four NVIDIA RTX 3090s. 256GB of ECC RAM. An AMD EPYC with 32 cores and 64 threads. Kubernetes. Jenkins. Open-source LLMs running locally via Ollama. Data going nowhere except to disk. A man named Ovais Quraishi has just finished writing a case study about it, for a product engineering team of six who needed to build Agentic-AI applications and did not have the million dollars that commercial cloud AI would have required. He called it a Private Cloud PaaS. He charged $26,000 for year one and $766 a month after that.

Nobody called Uber. Nobody called JPMorgan. Nobody forwarded the PDF to the GitHub Copilot procurement team or left it on a Shopify executive's desk with a sticky note reading read this before Q1 2026, please.

"Building the desired MVP using commercial cloud services would require over a million dollars - funding that isn't currently available."

- The 2024 case study, which was right there, available, just sitting there

Imagine, for a moment, the alternate timeline. It is early 2025. A VP of Engineering at Uber reads the case study. She forwards it up. Someone does the math. At their token volumes, Claude Sonnet API costs would run to tens of millions annually. A private inference cluster - AMD EPYC, a rack of GPUs, Kubernetes, the works - costs a rounding error by comparison. The data stays inside the building, which Legal has been asking about anyway. A purchase order is cut. By the time 2026 arrives and the per-token bills start arriving at other companies, Uber's AI budget is not a headline. It is just infrastructure. Boring, paid-off infrastructure.

In this timeline, that did not happen. In this timeline, Uber burned its entire 2026 AI budget in four months and is now rationing employee access. JPMorgan Chase - a bank, an institution built on not losing money - is watching AI token costs exceed some employees' salaries. GitHub Copilot users are discovering, via their first per-token invoice, exactly how much their coding assistant has been chatting. Shopify is wondering, quietly, if Sidekick can keep growing without the economics quietly inverting.

The Alternate History: What Investing in the 2024 Solution Woulda Looked Like

  • UberA private inference cluster at enterprise scale: call it $500k upfront. Woulda paid off in weeks at their token burn rate. Instead: entire annual AI budget, gone by April.
  • JPMorganAt 105M tokens/day, Claude Sonnet API runs $47,000/month. A private cloud runs $766/month for a team of six. Scale it up tenfold and it is still not eating anyone's salary.
  • ShopifySidekick on local inference: no per-query cloud bill, data stays on platform, scales with hardware not pricing tiers. Shoulda been the Sidekick architecture from day one.
  • GitHubCopilot on private inference for enterprise customers: zero per-token billing anxiety, predictable cost, genuinely unlimited usage. Woulda been a selling point, not a crisis.

The numbers from the 2024 case study are not subtle. At peak load - 105 million output tokens per day, ten users - Claude 3.5 Sonnet woulda run $1,575 a day, $47,000 a month, $564,000 a year. For output tokens only. Input tokens extra. Claude 3 Opus: $7,875 a day, $236,000 a month, $2.8 million a year. The private cloud: $766 a month, everything included, no rate limits, no overages, no calls to Legal when an agentic workflow goes wide. The case study noted, with admirable restraint, that the private cloud recoups its investment within a few months for almost all cases. The one exception is if you run barely any tokens at all, in which case you probably don't need an AI platform anyway.

$766 Private cloud
per month (2024)
$47k Sonnet API
per month at scale
$2.8M Opus API
per year at scale
$0 Rate limit fees
on private cloud

By 2026, Ovais himself has evolved the setup. The GPU server - too hot, too loud - was sold on eBay for a profit and replaced with a Mac Studio M4 Max, 128GB unified memory, silent, efficient, running 500 million tokens a week at $0.002 per million. Year two cost: $294. Total. The principle scaled down to one person running an operating system for his own life, generating more tokens weekly than most enterprise teams, paying less annually than a gym membership, beholden to no billing dashboard, no rate limit, no quarterly pricing review.

The blueprint existed. It was written down. It had line-item costs, architecture diagrams, a Kubernetes cluster layout, a Jenkins deploy time of under one minute, and a colocation bill of $450 a month. It was available in November 2024, when the AI bills of 2026 were still just a math problem waiting to be solved - or ignored.

"The meter was always running. The case study showed you exactly how much. Then you got the bill anyway."

- The bill, unimpressed

The coda: Ovais eventually subscribed to Ollama Cloud Pro. Twenty dollars a month, voluntarily, the day it launched, as a thank-you to the team that made his independence possible. He is paying for something he doesn't need, out of gratitude, while multinational corporations are paying for something they can't stop needing, out of desperation.

The playbook was written in 2024. The math was not complicated. The future was, in this specific and well-documented sense, entirely foreseeable.

You coulda called him. You woulda had a plan. You shoulda done the math.

The invoice is in the mail.