Hoonify

10/02/2026

We’re not actually taking Hoonify to the skies over Albuquerque… but with Balloon Fiesta here this weekend, we had to see what it would look like. 🎈

Grounded in New Mexico. Slightly less grounded in this post. 😉

Hoonify helps companies run AI and HPC they own, control, and trust, from our cloud to their air-gapped data center.

10/01/2026

Remember when your email app first started guessing the next word you'd type? That same technology just took on one of the hardest problems in math.

This month an AI lab announced it has solved Navier-Stokes, one of the $1 million Clay Millennium Prize problems. If the proof holds up, it'll go down in history.

Our HPC engineer noticed something else in the story. Part of the path to the proof ran through researchers who'd been using AI to work through their own ideas. A lot of us use AI that way now, so it's worth asking: when you share your best thinking with a model you don't control, who owns it?

Part 1 of our two-part series is live. 👇

An AI lab cracked Navier-Stokes using researchers' own ideas. Here's what that means for AI data ownership and your business.

10/01/2026

Live from Q Station! 🚀

Hoonify's CEO & Co-Founder, Andrew Clark, is on stage right now moderating "Funding Growth in Space Valley" at Space Valley Discovery Week. The panel covers how space and defense companies in New Mexico and the region find the funding to grow.

We're proud to call Albuquerque home and glad to help build the region's space economy.

09/30/2026

Where analytical functions end, the universe's beauty begins, and Monte Carlo lets us see it."

In 1946, Stan Ulam wanted to know his odds of winning at solitaire, so he pictured playing hundreds of random games and counting the wins. That idea became the Monte Carlo method. It gives you an answer and tells you how far off that answer might be. 80 years later it's used everywhere from physics to AI, and now quantum computing too.

This week, Los Alamos National Laboratory celebrated that legacy at MC80. Hoonify was proud to sponsor, and we were honored to celebrate with Stan's granddaughter, Rebecca Ulam Weiner.

As a New Mexico company started by national lab engineers, we're grateful to support the experts carrying this work forward.

09/24/2026

There's a claim we could make about our national laboratory work, and we don't make it.

We make simulations run faster. We do not make them more numerically accurate. Those are different things, and blurring them would be the easiest marketing move available to an infrastructure company.

Scientific validity sits with the codes and with the domain scientists who own their verification and validation programs. MCNP's benchmark suite is a Los Alamos process. It isn't ours and we don't speak for it.

What we are responsible for is making sure infrastructure change never outruns that discipline. Every kernel patch and platform update is exercised against the application's own regression suite before production, and the result goes into the software quality assurance record. If a change doesn't clear the bar, it doesn't ship, and the previously validated configuration stays.

Platform velocity is gated by the code's own V&V standard, rather than the other way around. Nobody should be asked to trust a faster number over a validated one.

Read our latest blog:

Most HPC is sold on speed. Sovereign HPC is judged on control and accreditation. Here is what that looks like in production.

09/17/2026

A cost comparison assumes both options are available to you. For a lot of organizations, one of them never was.

If your data can't cross a trust boundary, the serverless column isn't a cheaper alternative. It's not an alternative. Classified simulation data, export-controlled engineering documents, patient records, core IP. Some buyers are also barred from models of a given national origin regardless of licence or benchmark, which rules out the cheapest and most widely adopted options before price enters the conversation at all.

So the break-even number in our analysis, roughly 180 engineers running agents, is genuinely useful and completely beside the point for those readers.

Run it anyway. Not to decide, but to price the requirement. The gap between what you'll pay and what the same work would cost unconstrained is the premium you're paying for control, and it's a number worth carrying into the conversation rather than discovering afterwards.

Self-hosting an LLM vs API cost: the break-even is ~180 engineers running agents, plus the ops line nobody counts.

09/08/2026

Here's a 5 minute exercise that's worth more than most vendor evaluations.

Open your RAG pipeline configuration. Find the step that generates embeddings. Check where it runs.

If that's a hosted API, your documents crossed the boundary before the model ever saw them. And everything you built upstream, the on-prem cluster, the open weights you can hold, the network cable unplugged on purpose, was decorative for that pipeline.

This isn't a competence problem. Nobody skipped a review. Embeddings just don't feel like data the way documents do, so they don't get the same scrutiny. Caches and telemetry have the same problem.

Derived data counts. If it left, it left.

Go and check. We'd genuinely like to hear how many of you find something.

Sovereign AI means running AI on infrastructure, data, and models you control, under your own jurisdiction. A co-founder's plain-English definition, requirements, and architecture.

09/01/2026

18 months ago our biggest worry about AI was whether it worked. Now it's whether we can predict what it costs.

Frontier models from OpenAI and Anthropic are the best general-purpose reasoning engines available. Nobody's disputing that. But token pricing means your spend moves with customer growth, with usage patterns you don't control, and with model upgrades you didn't ask for. From a COO's chair, variable cost that scales faster than revenue is a margin risk, not an engineering line item.

Meanwhile the gap has closed for most of what companies actually ask AI to do. Classification, extraction, drafting, internal tooling, retrieval answers, routine coding help, first-pass support.

The other 20% still earns its premium. Complex reasoning where a wrong intermediate step compounds. Compliance-sensitive judgment. Anything where a bad output costs more than the extra tokens.

If you're about to scale an AI feature, read this before the invoice teaches you the same lesson:

Frontier models cost 5x to 100x more than open source alternatives. How we decide which workloads justify the premium, and how to forecast AI spend.

08/31/2026

The part of open weights that sounds scary and mostly isn't: licensing.

Most of the field is genuinely permissive. DeepSeek's V4 models and GLM-5.2 are MIT. Gemma 4, gpt-oss, Qwen3.8-27B, and Inkling Small are Apache-2.0. Nothing to track.

Two models carry revenue-gated licenses, and this is where people panic unnecessarily. Kimi K3 and Qwen's Max-class checkpoint only trigger their conditions if third parties reach the model AND your revenue crosses a line ($20M and $50M respectively). Self-hosting for your own engineers? Internal use is exempt under both. That covers most enterprise deployments.

Above the line, you need a separate agreement. That's a business development conversation, not a crisis.

Read the license file, not a summary. General information, not legal advice.

Worth the read before your next procurement conversation, along with the shortlist for each coding job:

Open-weight coding models had a remarkable year. Here are the strongest candidates for planning, hard debugging, and high-volume work - and how each compares to the closed models you already use.

08/28/2026

A question that comes up constantly: isn't sovereign AI just a government thing?

No. The term came out of national AI strategy, but the requirements apply to anyone whose data can't cross a trust boundary. Defense contractors and national labs, yes. Also hospitals, banks, law firms, and any company whose IP is the business.

Here's the part teams underestimate. The most valuable AI workloads are usually the ones you can't send to a third-party API. Classified simulation data. Export-controlled engineering documents. Patient records. Core IP. The data is valuable because it's protected.

And the quality gap keeps shrinking. Today's open-weight models match frontier models on a lot of production tasks and trail by low single digits on others. For work that can't leave your boundary, the real comparison was never open versus frontier anyway. It's open versus nothing.

Full post:

Sovereign AI means running AI on infrastructure, data, and models you control, under your own jurisdiction. A co-founder's plain-English definition, requirements, and architecture.