Remove the Ceiling.
Own the Infrastructure.

The One-Paragraph Version

Alpha is spending ~$10,000 per student per year on cloud AI — and it's the right call. You're building the best AI-powered school in the world and that takes real investment. But cloud AI has a ceiling: every token is rationed, every query goes through OpenAI or Anthropic servers, and the frontier models Joe actually cares about — the ones beating Claude Opus 4.6 on published benchmarks — aren't available via any API at any price. This proposal is about removing that ceiling entirely. For $2,500/month (month-to-month, hardware included, cancel anytime), Alpha gets a private AI cluster on campus: unlimited capacity, student data that never leaves the building, and access to frontier open-source models that outperform cloud AI on the benchmarks that matter.

Cloud AI Is Rationed. Sovereign Removes the Ceiling.

The $10,000/student/year you're investing in AI tools is impressive — and it should be. Alpha's bet on AI-powered education is the right one. But cloud AI infrastructure has a fundamental design constraint: it's shared, metered, and rationed.

• Token limits throttle the best use cases. Deep research, multi-step reasoning, long-context projects — these are exactly what Alpha students should be doing. Cloud API limits make that expensive to scale.
• Peak hours mean degraded performance. When 40 students are working simultaneously, cloud providers slow responses. You pay premium pricing and get shared infrastructure.
• The frontier models aren't available. Kimi K2.5 Thinking — which outperforms Claude Opus 4.6 on AIME 2025 — isn't on any API. DeepSeek R1 is available in limited form but not at the quality or capacity of running it locally. The best models don't have an "enterprise API" option.
• Student data leaves campus. Every query goes through OpenAI or Anthropic infrastructure. For a school with FERPA obligations, that's a real exposure — and a reputational one.

Sovereign installs a private cluster on campus. Your students get uncapped access to frontier models. Nothing leaves the building. The ceiling is gone.

Frontier Models. Not Available on Any Cloud.

The open-source model ecosystem caught up to — and in key areas, surpassed — cloud AI in 2025. These are the models running on the cluster:

Why Apple Silicon, Not NVIDIA

NVIDIA GPUs are the right choice for hyperscale cloud training. They are not the right choice for a 40-student campus cluster. Here's why the M5 Ultra is the correct hardware for this deployment:

• Unified memory architecture. A single M5 Ultra holds 512GB of memory on the same die as the compute — meaning a 400B parameter model loads in seconds and runs without the bandwidth bottlenecks that plague GPU server arrays.
• 2TB across 4 machines. Four M5 Ultras in a distributed inference cluster via the exo framework pool into 2TB. That number doesn't exist in the GPU world at this price point or power envelope.
• Power and heat are manageable. NVIDIA data center GPUs require dedicated power circuits and specialized cooling. The M5 Ultra cluster draws a fraction of that — it fits in a standard server rack in any room with A/C.
• The models that matter run on Apple Silicon. The largest frontier open-source models — 400B to 1T parameters — were designed to run in high-memory environments. The M5 Ultra unified memory architecture is ideal for exactly this class of model.

What Gets Installed

One rack. Turnkey. Sovereign handles everything from site survey to go-live.

Students open a browser tab. They see an interface that looks identical to ChatGPT — it's Open WebUI, an open-source frontend. They type. They get a response from a model that beats Claude Opus 4.6 on the math benchmark. They never know (or care) that the AI is running on a rack down the hall instead of in a data center in Virginia. Zero learning curve. Access from any device on campus.

Same system. Teachers get the same access. Custom agents for faculty workflows — lesson planning, grading assistance, curriculum research — can be built and deployed to specific user groups. Admin can audit usage, set access levels, and export logs at any time without involving a vendor.

The Questions Joe Will Ask

Pricing — Two Options

Both options include the same hardware, the same models, the same installation, and the same ongoing service. The difference is capital structure.

From Agreement to Live in 4–5 Weeks

Built by One of Your Students

My name is Kavin Lingham. I'm enrolled at Alpha School. I've spent the last year building Sovereign ATX — a private AI infrastructure company. In March 2026, I deployed the first external production installation at Black Sheep Coffee, a multi-location Austin café group. That deployment is running live.

I wrote the infrastructure configs. I rack the hardware personally. I manage deployments remotely via Tailscale. I wrote the service agreements and BAA templates. The whole operation runs out of my home lab on a Mac Studio M1 Ultra that I use for R&D.

Alpha School is where I learned to build things that actually work. This proposal is me offering Alpha School the same infrastructure I'm building for paying clients — as a founding client, at a price that makes it easy to say yes, with no lock-in because I want this to be a 5-year relationship, not a signed contract.

The pitch isn't "let me save you money." The pitch is: an Alpha student built the best AI infrastructure for education that currently exists, and he's offering it to Alpha first.

"Alpha gave me the time and freedom to build something real. This is me building it for Alpha."

Next Steps

This is a 30-minute conversation, not a 6-month procurement process. Here's how we move:

• 30-minute call — walk through any questions on the hardware, models, succession risk, student experience, or anything else
• Site walk (1 hour) — Kavin identifies rack location, power, and cable routing at the Alpha campus
• Agreement signed — one-page month-to-month service agreement; hardware order placed same day
• 4–5 weeks to live — 40 students with unlimited access to frontier AI, all on campus