Putting the open back in OpenAI

Welcome to Runtime! Today: OpenAI releases two open-weight models ahead of the expected launch of GPT-5, why China was already deeply involved in Microsoft SharePoint before last month's security fiasco, and the latest funding rounds in enterprise tech.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Token gesture

In recent years, OpenAI's name has been a little misleading; it hasn't shipped any major large-language models that could claim a degree of openness since 2019's GPT-2. But that changed on Tuesday with the release of two new open-weight models that the company said are about as powerful as some of its leading closed models.

OpenAI announced gpt-oss-120b and gpt-oss-20b Tuesday, comparing them to its o4-mini and o3-mini models, respectively. The text-only models can be run anywhere customers like for much cheaper than the more performant models in OpenAI's arsenal, and they offer users far more insight into their inner workings than other OpenAI models.

• "Open" is a loaded term in the AI model context, and OpenAI was careful to refer to the new models as "open weight," which means developers can examine how the models make decisions about their output.
• The code behind the new models is also available under the permissive Apache 2.0 license, unlike other open-weight models like Meta's Llama that place restrictions on how the model can be used.
• However, the training data behind the gpt-oss models remains under lock and key, which open-source AI experts such as AI2 believe is an obstacle to truly understanding how AI models perform.

But this is still a shift in OpenAI's product strategy, which revolves around charging consumers and enterprises for access to the models it has spent billions training over the last several years. The new models arrive a little over six months after China's DeepSeek caused a mini-panic in AI circles with the release of an open-source model that was nearly as good as anything that had been released until that point.

• It's no secret that while OpenAI has captivated the public with its GPT models and ChatGPT service, enterprises prefer models from its rival Anthropic (which also released a new version of its Claude model Tuesday) when building generative AI apps.
• And according to VentureBeat, OpenAI President Greg Brockman said during a press conference that most of OpenAI's paid customers are using a mixture of its closed models and open-source models from other providers.
• But the gpt-oss models will allow OpenAI to work with a wider variety of enterprise customers, especially ones that were hesitant about sharing corporate data with the company.
• That's because the models can be run and customized locally, on anything from a laptop to a server to a cloud instance controlled by the user.

The new open models arrived later than originally expected, which has been par for the course among AI model providers over the last year as it becomes harder to make substantial improvements in model performance over time. But as AI developers await the near-imminent launch of GPT-5 — which will not be cheap — OpenAI's new gpt-oss models arrive right as companies are starting to wonder if they really need all the bells and whistles provided by the frontier models.

• Some applications will always need the best model their developers can afford, but attention is turning to small-language models that were trained through a much cheaper process and are probably good enough for a lot of applications.
• The new models slide right into that conversation, and because they're available under the Apache 2.0 license they could expose OpenAI to a new set of potential customers.
• And perhaps no company was more excited about the launch than AWS, which finally gets to say that it can offer OpenAI's models to its customers thanks to that permissive license.

Thanks for reading and supporting Runtime! As we head into the busy fall launch season, please consider making Runtime a part of your marketing strategy over the rest of 2025. We rolled out several new sponsorship opportunities and pricing packages this summer, and you can find more details about those options here.

Oversharing

As the SharePoint debacle enters the ransomware stage of the exploit, we're learning more about how Microsoft maintained the on-premises version of SharePoint that was compromised by hackers believed to be working on behalf of the Chinese government. After ProPublica reported several weeks ago that Microsoft used staff based in China to maintain cloud infrastructure for the U.S. government, it turns out that wasn't the only group based in China working on a key piece of its software portfolio.

According to a new ProPublica report, engineers based in China have been responsible for fixing bugs in the on-premises version of SharePoint that was compromised in the attacks. Microsoft said that the team in question was "supervised by a US-based engineer and subject to all security requirements and manager code review," and that "work is already underway to shift this work to another location.”

Last month security experts suggested that one of its security partners might have leaked details of its proposed patch for fixing the SharePoint vulnerabilities at issue, but the new report suggests another possibility. While there's no evidence that the SharePoint support workers were involved, ProPublica reported that there have been ongoing concerns about U.S. companies relying on developers based in China because "experts say it is difficult for any Chinese citizen or company to meaningfully resist a direct request from security forces or law enforcement."

Enterprise funding

Observe raised $156 million in Series C funding for its observability software, which is based around the evolution of the data lake and agentic AI.

Anaconda landed $150 million in Series C funding as it builds out its AI development platform, which helps developers find packages of Python code they can trust.

Fal scored $125 million in Series C funding for its AI compute infrastructure and developer platform, which helps companies generate all those terrible AI images and videos currently polluting the internet.

Oxide raised $100 million in Series B funding as it continues to develop a design for on-premises servers that are just as good as the ones designed by the hyperscale cloud providers.

Noma Security scored $100 million in Series B funding for its AI security platform, which helps customers understand the risks posed by their AI agents and close the loopholes.

Pantomath landed $30 million in Series B funding as it continues to build out its data observability platform for AI app development.

The Runtime roundup

Super Micro's stock tanked after it reported revenue and earnings below Wall Street's expectations and lowered its revenue forecast for its fiscal year.

AMD beat estimates for quarterly revenue but missed on profit, and even though it raised its revenue guidance for the current quarter the day traders were not pleased.

Thanks for reading — see you Thursday!