In 1968, Robbie Robertson, bereft of ideas for his next song, looked down at the sound hole of his 1951 Martin D-28 and saw a stamp noting the guitar maker’s factory  “Nazareth, Pennsylvania.”

Inspiration struck. He wrote: “I pulled in to Nazareth, was feeling 'bout half past dead…” and called the song “The Weight”.

Weights are suddenly at the center of the discussion of Artificial Intelligence models this week. And, like the song’s protagonist, the decision of how to handle all these weights – or whether to leave them all behind – is the heaviest of matters. 

On one side you have the fantastic success of closed weight models:  OpenAI's GPT-5.5, Anthropic's Claude and Google's (GOOGL: NASDAQ) Gemini. These large language models lock down the ability to adjust their reasoning, or “weights”. And in the last few weeks we’ve seen a surge of interest in adjustable weight models, including Meta's  (META: NASDAQ) Llama, DeepSeek's V4 line, Mistral, Alibaba's (BABA: NYSE)  open Qwen family and especially Z.ai's GLM 5.2.

The battle over these opposing technology models may well decide who will emerge as the dominant AI firm — and, indeed, whether the US or China wins at AI.

A model's “weights” are its trained parameters — the interpretations of the billions of numbers the model learned during training. The weights are the model. 

How to understand this? Think of requesting a song. You might want to hear a 12/8 groove, a female vocalist, bass, drums, guitar, a synthesizer – all wrapped up as  a cute pop song.

A closed model would look at that request and offer up the finished recording. It might be Aretha Franklin's "Rock Steady" or Olivia Dean's "Be the Man I Need." You can pick which finished version you want, but you never touch the faders. The mix was set by a weight you can’t adjust.

An open-weight model hands you the session itself. Every instrument arrives on its own track, the whole multitrack on a drive you can carry into your own studio. Turn the bass up. Mute the synthesizer. Re-cut the vocal in another language for another market, run the drums through your own console and release what you make without telling the original studio a thing.

Why would a user want open weights?

The task only open weights can do. A hospital system that wants a model reading clinical notes hits a wall no closed API can climb: protected health information legally cannot leave infrastructure the hospital controls, and self-hosting requires weights you can download. Meta publishes a deployment guide for running Llama inside a HIPAA boundary because the demand is that concrete. The same requirements exists in defense, parts of banking and any strict data-residency jurisdiction.

The advantage only closed weights deliver. During Fable 5's testing, Stripe ran a single codebase-wide migration across its 50-million-line Ruby codebase in one day — work the company said would have taken a full team more than two months. Stripe didn't want to adjust every part of the model. It had a job it wanted done, now. That's most of the market: no compliance wall forcing the work in-house, no need to retune anything — just a hard task and a deadline, which makes the best available model the whole decision. The session files are for customers who need control. The finished record is for customers who need results.

Three forces turned a philosophical debate into a procurement decision.

The first is capability. Artificial Analysis grades models on a common battery of tests to its Intelligence Index. In early 2025 the best open-weight model scored 13 points below the best closed one.

But the gap between closed and open is closing. GLM-5.2, the Chinese open-weight leader released in June, scores 51 against Claude Fable 5's 60 and GPT-5.5's 55. On some individual tests the gap is already gone: DeepSeek V4 Pro fixes real software bugs as well as GPT-5.5 does, at a fraction of the price, under a license anyone can use commercially.

The second is cost. Open-weight models are far cheaper. DeepSeek V4 Flash handles tasks across all six of Artificial Analysis's new industry capability indices for less than $0.04 per task. Claude Fable 5 might be the smartest model money can rent — and at $3.25 per task that closed model was the most expensive.

To address this Fastino, a hot Palo Alto startup founded in 2024, built its entire business on open weighted small language models. Its Pioneer tool, launched in April, lets a developer take an open model — Qwen, Gemma, Llama — and retune it for a single job with one written prompt: pulling account numbers from documents, redacting personal data, sorting text into categories. None of that is possible with a closed model, because retuning requires the weights, and closed labs don't hand them over. Fastino's own open models have been downloaded more than 6 million times, and Khosla Ventures, Insight Partners and Microsoft's venture arm have put roughly $25 million behind the bet. 

The third is government, and here the ground is still moving. 

Europe rewards openness. The EU AI Act, enforceable Aug. 2, imposes new documentation and transparency duties on model makers — but grants openly licensed models a partial exemption. The logic: a model anyone can download and inspect is already transparent in a way a black-box API never is, so the law demands less of it. In Brussels, publishing your weights buys regulatory relief.

Washington currently has no standing rules for model weights at all — the Biden administration's framework, which controlled closed frontier weights and exempted open ones, was tossed out in 2025. The Trump Administration repealed that plan, but hasn’t replaced it with anything clear. 

In late June a hasty-crafted executive order asked labs to submit frontier models for pre-release review, after a demand to pull Claude Fable 5 and Mythos 5 offline for 19 days and a subsequent customer-by-customer approval rollout for OpenAI's newest model. Each new restriction on a US closed model makes the un-gatable open alternatives, increasingly Chinese ones, a little more attractive.

With Anthropic and OpenAI IPOs looming the weights issue may inform investors as to what they’re buying. An OpenAI investor is buying both sides of the bet. The company gives away its gpt-oss models to keep the customers who must run AI on their own computers — hospitals, banks, governments — from drifting to open rivals, while charging for its best models, where the money is. The risk is that the free tier undercuts the paid one: every capable model OpenAI gives away teaches customers that good AI should cost little, and makes it harder to charge for anything but the very best. Running both sides at once is hard, and no company has done it well for long.

An Anthropic investor is buying one side only. The company has never released weights, and its entire value rests on a single proposition: that its models will stay far enough ahead that customers keep paying the highest prices in the industry. June showed both faces of that bet. When the government ordered Fable 5 offline, it could comply within hours — because the weights had never left its servers, there was exactly one switch to flip, off and then back on.

We’ll see if they can handle the weight. 

“My bag is sinkin' low and I do believe it's time.

To get back to Miss Annie, you know she's the only one.

Who sent me here with her regards for everyone.”

– The Band

NewsNation’s Alicia Nieves took on the tough topic of propaganda from China, Iran and Russia seeking to inflame the already-raging debate of AI Data Centers. I took the tack that the debate is real, but the inflaming doesn’t have to happen.

SEE THE NEWSNATION HIT HERE →

The unlock for agentic workflows may turn out to be an unglamorous one: cross-border payments. That's the thread I picked up in Aaron Ricadela's new eBook, "Battle Is On for $2.5 Trillion Cross-Border Payments Market," where I argued that moving money across borders becomes a serious accelerant the moment agents start transacting on their own behalf. The conversation pulled me straight back to my fintech days at Ripple and Braintrust, when "settle it instantly, anywhere" was the whole ballgame — and payments and agents were nowhere near colliding the way they are now.

SEE THE FREE EBOOK HERE →

Bubble or no Bubble? The signal everyone's watching for the AI top — spare compute changing hands — is the wrong one. Asked by the InvestorsObserver newsletter what would first warn that AI spending is cooling, I pushed back on reading deals like xAI renting idle capacity to Anthropic as the beginning of the end: that's not the buildout cracking, it's a verdict on Grok, and the freed-up capacity is exactly what hands a challenger like Cursor its shot. The real tell is the day AI usage flattens — and we're nowhere near it.

SEE THE PIECE HERE →

On Wednesday afternoon last week San Francisco recored the coldest temperature in America (60º F at SFO) — so why would I leave? I’ll be in our San Francisco office all week so text or call and I’ll do my best to captivate you on this amazing moment in technology.

Written reports are available to clients, our IPO reports for $5,000. Video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.

^{We certify that (1) the views expressed in this report accurately reflect our views about all of the subject companies and securities and (2) no part of my compensation was, is or will be directly related to the specific recommendations or views expressed in this report.}

^{Important disclosures}

^{Important disclosures are available by calling (347) 619-2489 or writing to Epistrophy Capital Research, One Ferry Building, Suite 201, San Francisco, CA 94105.}

^{Epistrophy Capital Research is an independent research provider and does not operate as a financial institution. Epistrophy explicitly does not provide investment advice, stock recommendations, or solicitations to buy or sell any securities.}

^{The research reports provided by Epistrophy Capital Research contain opinions derived from publicly available information, issuer communications, recognized statistical services, and other reputable sources considered reliable. However, Epistrophy does not independently verify the accuracy or completeness of such information and explicitly disclaims responsibility for any errors or omissions.}

^{Opinions and analysis contained within Epistrophy's research reports are current only at the time of publication and are subject to change without notice. Readers must independently verify facts and conduct their own due diligence before making investment decisions.}

^{Epistrophy Capital Research does not consider or evaluate individual investor circumstances, including investment objectives, financial situations, or risk tolerance. Investing in securities, particularly small-cap and micro-cap stocks, involves substantial risks, including significant volatility and potential loss of principal. Readers are strongly advised to consult their financial advisor or another qualified professional before acting on any information provided. Readers should assume Epistrophy Capital Research, its principals or its contributors may have positions, long or short, any of the companies discussed and the Epistrophy Capital Research principals or contributors may have had or currently have business interests in the companies discussed.}

^{Past performance referenced in Epistrophy reports is not indicative of future results. Security prices can fluctuate widely, and investors should be aware that investments can result in significant financial losses. Epistrophy Capital Research or its Unless explicitly stated otherwise, prices quoted in reports reflect market closing prices from the previous trading day.}

^{Epistrophy Capital Research publications are intended solely for direct recipients and should not be redistributed or shared with third parties without explicit permission from Epistrophy Capital Research LLC.}

^{Epistrophy Capital Research reports are provided strictly for informational purposes and do not constitute a comprehensive analysis of any company, security, or industry. No content within these reports should be considered accounting, tax, legal, or professional advice.}

^{Links or references to third-party websites or external resources are provided solely for informational convenience. Epistrophy Capital Research expressly disclaims endorsement of, and responsibility for, the content, accuracy, or reliability of such external information. Accessing third-party information is done entirely at the user's own risk.}

^{For additional details, clarification, or specific inquiries regarding Epistrophy Capital Research reports, please contact Epistrophy Capital Research LLC directly.}