Microsoft just took sides in AI policy fight | Official Website of Louis Velazquez, entrepreneur, finance guy and tech innovator

Every powerful technology eventually reaches the same argument. Somebody has to decide whether the thing gets handed out or locked up.

Encryption had that fight in the 1990s, when the federal government classified strong cryptography as a munition and tried to keep it away from ordinary people. Software had it a decade earlier, when a small group of programmers insisted that code should be readable by anyone willing to read it.

That second fight is the one worth remembering, because the side that looked reckless at the time ended up running the century. Open source software now underpins most of the internet, along with the systems that federal agencies and the U.S. military rely on every day. Nobody calls that reckless anymore.

Artificial intelligence has arrived at the same intersection, and the stakes are bigger this time, because AI is not merely software. It is the thing that writes the software.

For three years, America’s largest AI labs have argued that their most capable models are too dangerous to hand out. Washington has been broadly sympathetic. Chinese labs have not been listening.

On July 24, Microsoft (MSFT) stopped hedging and put its name at the top of a document that says the opposite.

Washington leaned toward restricting downloadable AI models. Microsoft’s July 24 letter argues the opposite.

d3sign / Getty Images

Why open weight AI models became a Washington problem

An open weight model is one anyone can download, inspect, modify and run on their own hardware. The company that trained it gives up control the moment the file goes public. A closed model stays behind an application programming interface, or API, where the developer decides who gets access and at what price.

More Artificial Intelligence:

For most of the modern AI boom, the best models were closed and American. That arrangement quietly held U.S. policy together.

It stopped holding on July 16, when Beijing-based Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model it planned to release freely. One widely read estimate now puts the lag between open models and the closed frontier at roughly “3-5 months,” according to Interconnects, down from the six to nine months the field had been assuming.

Washington noticed. The Commerce Department had already spent June testing how far its authority reaches, imposing a license requirement on the distribution of two Anthropic frontier models before largely withdrawing it two weeks later, according to Export Compliance Daily.

That episode showed the industry something uncomfortable. The government is willing to treat a model file the way it treats a weapons component.

What Microsoft signed and who signed alongside it

The letter is titled “Open Weights and American AI Leadership,” and it lives on Microsoft’s own corporate responsibility site rather than a trade group’s. That placement is the tell. Microsoft is not a co-signer here so much as a host.

The argument runs against the industry’s own safety orthodoxy. “Relying solely on closed models is not inherently safe,” the letter says, according to Microsoft, which goes on to argue that concentrating capability behind a few providers creates single points of failure.

I counted 35 names on that page on the morning of July 26. The coverage on July 24reported 25.

Here is how the list moved:

Twenty-five companies had signed when the letter published on July 24, including Nvidia (NVDA), Meta (META), IBM (IBM), Dell (DELL) and Palantir (PLTR), according to CNBC.
OpenAI was absent at launch, according to Tom’s Hardware.
OpenAI now appears among the signatories, alongside later additions including Cisco (CSCO), Box (BOX), DoorDash (DASH) and GitHub, according to Microsoft.
Elon Musk backed the letter publicly without SpaceX (SPCX) signing it, calling it something he gave his “full support,” according to CNBC.
Anthropic and Alphabet’s (GOOGL) Google remain off the list.

At publication, “none of the signatories sells access to a closed frontier model,” according to Tom’s Hardware. OpenAI’s later addition complicates that reading, and it is the single most interesting thing about the document.

What the open weights fight means for Microsoft stock

Here is where this stops being a policy story and starts being a portfolio one.

Microsoft carries a market capitalization of roughly $2.84 trillion, down about 25% over the past year, according to StockAnalysis. If you hold an S&P 500 index fund in a 401(k), you own a piece of this argument whether you followed it or not.

The commercial logic is not subtle. Microsoft sells cloud capacity. Every model that runs on customer-controlled infrastructure is compute Azure can rent, and Microsoft does not pay a licensing toll on weights it did not train.

Closed frontier models cut the other way. They concentrate margin at the lab, which is why rivals keep building their own silicon to escape the same math, and Microsoft has been paying that toll to a partner it also competes with, as I reported for TheStreet in June.

Competitors see the same incentive. U.S. labs are “clearly worried,” Mozilla chief technology officer Raffi Krikorian told Axios, arguing that executives would not lobby against open weights unless they saw a genuine competitive threat.

What struck me in my analysis of the signatory list is that almost nobody on it sells intelligence directly. They sell chips, clouds, tooling, distribution and security. Cheap, abundant, downloadable models make all five of those businesses larger.

That does not make the argument wrong. It does mean the reader should price it as advocacy rather than testimony.

Where the open weights fight heads next

Three things worth watching, none of them abstract.

Moonshot said it would publish K3’s weights on July 27, according to Axios. That would put a near-frontier Chinese model on American laptops and force Washington to respond to a fact rather than a forecast.

Microsoft reports fiscal fourth-quarter results on July 29. Azure growth has been the number that moves the stock, and a strategy built on renting compute for models Microsoft did not train is now company policy in writing.

And Congress is still working through the AI Kill Switch Act, legislation that would require a shutdown mechanism for AI models. A model whose weights sit on 40,000 hard drives does not have a switch to flip.

The signatories know that. They wrote the letter anyway, which tells you they would rather argue about competition now than compliance later.

For everyone else, the practical read is simpler. The cost of using capable AI is heading down faster than the labs charging for it would prefer, and the companies betting on that outcome just told Washington so in public.