By Vaughn Woods, CFP®, MBA
Vaughn Woods Financial Group, Inc. | August 2026
There is a vocabulary problem in artificial intelligence right now, and it is costing people money.
Not because the words are difficult. Because the words are doing work that most investors never inspect. When a headline says a company released a “frontier model” or made its weights “open,” a picture forms in the reader’s mind — usually a picture of dominance, or of surrender. Neither is typically accurate. And decisions get made on top of that picture.
I want to take three terms apart. Then I want to show you why the structure underneath them is more interesting than the horse race everyone is covering.
Foundation models
Start with the broadest of the three, because it is the one most often used as a synonym for the other two.
A foundation model is a large model trained once on an enormous sweep of data, which can then be cheaply adapted to many specific tasks rather than rebuilt for each one. Every modern large language model is a foundation model. Not every foundation model handles language — the category also covers image, audio, and video systems built the same way.
The etymology is the interesting part, and I think it has been almost entirely lost in usage.
Stanford’s Center for Research on Foundation Models coined the term in a 2021 report, after what one of the authors described as several weeks of internal debate and an explicit tournament among candidate names. They chose “foundation” over “foundational” on purpose. “Foundational” would have implied these systems supply fundamental principles. “Foundation” was meant to convey something narrower and more honest: a foundation is built first, it is load-bearing, and it is by definition, unfinished. It requires substantial subsequent building before it is useful for anything.
The researchers who named the category were describing incompleteness. The market heard permanence.
That gap between what a word was built to mean and what it is heard to mean is the subject of this article.
Open weights
An AI model is, at bottom, an enormous collection of numbers that the system learned by processing text. Those numbers are called the weights. They are the model.
Some companies keep their weights private. You can send a question to the model over the internet and receive an answer, but you never possess the thing itself. That is a closed model, and it is the arrangement most people are familiar with.
Other companies publish the weights. You download them, run them on hardware you control, modify them, and deploy them however you like. Those are open-weight models.
The distinction that trips people up: open weights is not the same as open source. The parameters may be public while the training data and the full code remain private. You are given the finished cake, not the recipe. There are exceptions — NVIDIA publishes weights, training data, and training recipes together for its Nemotron family — but they are exceptions.
Frontier models
A frontier model is simply whichever model sits at the leading edge of capability at a given moment.
There is no certifying body. No trophy. The Frontier Model Forum’s working definition centers on general-purpose systems trained with enormous computational budgets that exceed the state of the art across multiple domains, and the EU AI Act uses a somewhat lower threshold to trigger regulatory obligations. But in practice the label is relative and it decays. A model that was frontier eighteen months ago is an ordinary model today.
Hold onto that. It is the whole point of this article.
The cloud layer
Almost nobody runs these models on their own hardware. The models are too large and the hardware is too specialized. So, the cloud providers host them.
Amazon’s Bedrock service is a useful illustration. It presents dozens of models — from Google, Mistral, OpenAI, NVIDIA, Alibaba, and others — behind a single interface, so a business can switch from one to another without rebuilding anything. Amazon added eighteen open-weight models in a single announcement in December 2025 and has continued adding them since.
Notice what Amazon is doing. It is not betting on which model wins. It is making the models interchangeable, and selling the electricity.
Where things actually stand
As of early August 2026, Anthropic’s Claude line leads most independent measurements. On the Artificial Analysis Intelligence Index, a composite of nine difficult evaluations, Anthropic holds the top position. On the benchmark designed to measure performance on real professional work rather than puzzles, Anthropic holds the top three positions outright. Blind human preference voting on LMArena tells a consistent story.
Google’s Gemini is strong on scientific reasoning and long documents. xAI’s Grok is the least expensive at the frontier and the best at knowing what happened this morning.
Now here is the part worth writing down: that ranking has changed repeatedly in the past year, and it will change again. The prediction markets that price this question in real time reprice it constantly.
The pattern underneath
The layer that appears most impressive is not necessarily the layer that captures the value.
NVIDIA gives away frontier-scale models — weights, data, and recipes — because every organization that fine-tunes and deploys one of those models buys NVIDIA hardware to do it. The model is not the product. It is the demand generator for the product.
Amazon commoditizes the model layer for the same structural reason. If every model is one API call away from every other model, no single model developer can extract much of a premium, and the margin settles into the infrastructure beneath them.
Meanwhile the model developers themselves are engaged in enormously expensive competition to hold a lead that historically lasts a few months.
I am not going to tell you what to conclude about any specific security, and nothing here is a recommendation. But I would observe that “who has the best model” and “who earns durable returns” are separate questions, and the financial press has a strong tendency to conflate them.
Why I think about this in a fiduciary frame
My work rests on a conviction that investment decisions are neurological acts of context-building. We do not evaluate facts in isolation. We evaluate them against a frame, and the frame is usually supplied to us by whoever spoke last.
The AI narrative is an unusually clean demonstration. A reader encounters “frontier model,” forms an impression of permanent superiority, and quietly attaches that impression to a stock. But the term describes a temporary position in a race, not a moat. The frame did work the reader never authorized.
The Stanford researchers picked “foundation” precisely because a foundation is unfinished. Somewhere between the report and the earnings call, the word stopped meaning “the beginning of something” and started meaning “the bedrock of everything.” Nobody decided that. It simply drifted, the way words do, in the direction of whatever the listener already wanted to believe.
This is the same mechanism that produces concentrated positions in whatever performed well most recently, and the same mechanism that produces liquidation at the bottom of a decline. Not a failure of intelligence. A failure of context.
The practical response is not to become an expert in machine learning architecture. It is to build a decision structure that does not require you to correctly predict which company leads a fast-moving technology race — because that prediction is unreliable and always has been. Diversification is not a hedge against ignorance. It is an acknowledgment that leadership rotates faster than conviction does.
The companies quietly winning the AI buildout figured this out already. They positioned themselves so they do not have to know who wins. That strikes me as a reasonable posture for a portfolio, too.
If your allocation has drifted toward a story you have not examined recently, that is worth a conversation. Not urgency — examination. There is a meaningful difference, and my job is to help you hold onto it.
References:
Stanford HAI, “Reflections on Foundation Models,” October 2021. https://hai.stanford.edu/news/reflections-foundation-models
“Foundation model,” Wikipedia, accessed August 2026.
digital-m, “Frontier model : c’est quoi exactement,” March 2026.
NVIDIA, “Nemotron AI Models,” developer documentation.
DataCamp, “Frontier Models Explained,” January 2026.
FindSkill, “What Is a Frontier Model? Plain-Language Guide,” June 2026.
Amazon Web Services, “Amazon Bedrock adds 18 fully managed open weight models,” December 2025.
BenchLM, “Artificial Analysis Intelligence Index Leaderboard,” August 2026.
Artificial Analysis, “GDPval-AA v2 Leaderboard,” August 2026.
ADSLZone, “Chatbot Arena top mundial de modelos IA,” July 2026.
Tech Insider, “Grok 4.5 vs GPT-5.6 vs Gemini 3.1 Pro,” July 2026.
Polymarket, “Which company has best AI model end of July?” July 2026.
NVIDIA, “NVIDIA Debuts Nemotron 3 Family of Open Models,” December 2025.
Disclosures:
Vaughn Woods, CFP®, MBA is President and Founder of Vaughn Woods Financial Group, Inc., an Investment Advisor Representative of Bolton Global Capital, Inc. Client assets are held in custody through Pershing LLC, a subsidiary of Bank of New York Mellon. This article is for informational purposes only and does not constitute personalized investment or tax advice.
We are unable to accept orders via email. If you wish to place an order, please consult your registered representative or contact the home office trading desk at (800) 649-4554.
This email system is for business purposes only and any information, including attachments, transmitted in this email is not confidential. Any message may be reviewed by authorized compliance personnel and/or produced to regulatory agencies or others with a legal right to access such information.
Past investment performance is not indicative of future results. Securities offered through Bolton Global Capital, Inc., Bolton, MA. Member FINRA, SIPC. Advisory services offered through Bolton Global Asset Management, a registered investment advisor, 579 Main St., Bolton, MA 01740 (978) 779-5361.
Investors should be aware that there are risks inherent in all investments such as fluctuations in investment principal. Past performance is not a guarantee of future results. Asset allocation cannot assure a profit nor protect against loss. Although the information has been gathered from sources believed to be reliable, it cannot be guaranteed. Views expressed in this newsletter are those of Vaughn Woods and Vaughn Woods Financial Group and may not reflect the views of Bolton Global Capital or Bolton Global Asset Management. The information provided is for general informational purposes only and should not be considered individual recommendation or personalized investment advice. Representatives and Advisors of Vaughn Woods Financial Group are not tax or legal professionals, if you need tax or legal advice, please make sure to consult a tax professional/CPA and/or a lawyer. VW1/VWA0404