BIG TECH

Exclusive: OpenAI unveils protocol to stretch compute

OpenAI is getting creative to deal with the industry's imminent compute crunch. 

On Wednesday, the ChatGPT maker and a coalition of researchers from AMD, Broadcom, Intel, Microsoft, and Nvidia published a paper offering a rare look into company’s training stack, debuting a new compute networking protocol designed to make GPU clusters faster, more reliable and conserve precious compute cycles. 

The protocol, which has been in the works for two years, is instrumental in OpenAI scaling the compute that it needs to continue building bigger and better models, noting in a blog post that the networking approach accelerates its vision for Stargate, the company’s long-term effort to garner the compute it needs to build and scale cutting-edge AI. 

The paper introduces a protocol called MRC, or Multipath Reliable Connection, which essentially tackles two main issues with the networks that serve as the connective tissue of AI infrastructure: Congestion and failures, Mark Handley, OpenAI’s networking lead, told me. As GPU clusters grow, these are problems that become more arduous to solve. 

• This protocol relies on “packet spraying,” said Handley, which essentially scatters data along hundreds of paths in the network simultaneously to prevent any one network link from getting congested. This also reduces the amount of “tiers” in a GPU cluster, resulting in “flatter” networks that use up less of the data center’s compute and power. 

• When handling network failures, MRC detects and reroutes when paths go down in microseconds. This allows GPU clusters to continue training seamlessly, even if parts of the network break down. 

• Additionally, the MRC pairs with a protocol called SRv6, or IPv6 Segment Routing, which essentially tells data the exact path it needs to take through a network, rather than forcing the network switches to do the routing work themselves, further reducing the energy requirements of these switches and the data center more broadly. 

“We want to use as much compute as we can get, but also we want to make sure that we're using it efficiently and effectively, and this is a critical component of that,” Greg Steinbrecher, OpenAI’s workload lead, told The Deep View in an exclusive interview. 

The protocol is already in use in OpenAI and Microsoft’s largest training clusters, including the Oracle site in Abilene, Texas, and in Microsoft’s Fairwater supercomputers, and has been used to train multiple OpenAI models. 

When implemented, this new protocol introduces several major downstream advantages, Steinbrecher told me. Conventional, large-scale AI training jobs are a “failure amplifier” for GPU clusters, he said: If one thing goes wrong, the ripple effect forces the process to grind to a halt, leaving GPUs to sit idle. Network congestion, additionally, slows the rate at which researchers can innovate. 

MRC circumvents these issues, allowing OpenAI to “turn the crank on our entire research pipeline much faster,” Steinbrecher said. “That allows us to make better use of the resources that we have.” 

The MRC specification is available through the Open Compute Project under an open license. Steinbrecher emphasized the importance of this, claiming that this protocol is not one in which OpenAI is trying to “differentiate,” but rather move the entire industry past what they consider a legacy bottleneck. Handley said that the infrastructure industry has reached a point where it’s worth establishing open standards, “as opposed to each of these large companies doing their own thing.” 

“Several players in the industry have their own in-house implementations of protocols … that type of market fragmentation is bad for the networking industry,” Steinbrecher said. “You want everyone's energy going in one direction and pushing together, and then everyone moves faster as a result.”

At the end of the day, all roads lead to compute. The math is simple: The more efficient data centers and AI factories can be, the more capacity OpenAI has to train its models. And by making the protocol an open standard, the company stands to deepen the well of compute resources for the entire industry. In turn, that stands to benefit OpenAI. With MRC, OpenAI wants to unleash a rising tide that lifts all boats, which is increasingly critical as AI companies stare down a looming compute shortage.

TOGETHER WITH SCRIBE

The Secret AI Platform Of Fortune 500 Companies

Nearly half of all Fortune 500 companies are using Scribe — and it's easy to see why.

Scribe is a workflow AI platform that gives leaders visibility into how work actually gets done, then makes it effortless to document and scale what’s working best.

And the results? Well, they speak for themselves:

• HubSpot cut time spent responding to a customer by 93%

• New York Life employees save up to 41.6 hours per month on documentation

• Mindbody reduced time spent creating customer training content by 40%

If you're looking to increase operational efficiency, improve workflows, and boost your ROI, learn how companies are already doing it with Scribe — and schedule your demo right here.

RESEARCH

Quantum, AI spur biological breakthroughs

For years, IBM has been at the forefront of quantum computing. At this year's Think conference, that ambition was on full display.

During the opening keynote, IBM CEO and Chairman Arvind Krishna highlighted quantum’s potential with the technology having the capability to unlock new discoveries at an incredibly quick pace — including AI developments.

“Quantum can help uncover what AI cannot yet compute, then AI learns from the quantum and can make faster and faster progress on algorithms and on computations to give you a state of where we are,” said Krishna. 

To showcase the tangible use cases of quantum computing, IBM highlighted a biological research milestone achieved with the Cleveland Clinic and Riken: using two of the IBM quantum computers and two of the world’s most powerful supercomputers, the companies were able to simulate protein complexes spanning up to 12,635 atoms. In October 2024, it was only able to simulate 10. 

• This is important, as the molecules in your body are proteins, or the “workhorse in the cell” that allow people to exist every day, as Serpil Erzurum, EVP and chief research and academic officer at the Cleveland Clinic’s Lerner Research Institute, explained during the keynote. 

• Understanding the 3D structure and motion of a protein is key in biological research, as it helps researchers understand how a drug candidate could bind to a protein and develop effective drugs. Yet, it has remained a challenge as classical computers can only approximate solutions. Erzurum emphasizes that this development is “a moment.”

“Everyone will want to see what these structures look like to understand biology, disease, what's going wrong if it's not working, and more importantly, what can I make to fit into the three dimensional structure, to change the structure of that protein–because that's therapy, and that can make a difference in life,” said Erzurum. 

Another example Erzurum noted is using quantum computing and machine learning to dramatically speed up the identification of which treatments a harmful microbe is sensitive to, potentially saving lives given that infections remain a leading cause of death globally. 

In a separate Q&A with analysts and select press, Krishna did make it clear that in the next three years, he does not see quantum as replacing either AI or classic CPUs, but rather it will solve problems the two cannot solve, such as the modeling molecules example.

While quantum computing ranks among the most cutting-edge technologies available today, so advanced that it can be difficult to fully realize, it is already demonstrating tangible results. Its relevance is also becoming increasingly difficult to overlook, particularly given its potential to address some of the most pressing challenges facing the AI industry, including the growing demand for compute power. That said, it is important to contextualize the technology's current trajectory: Widespread mainstream adoption and full commercial deployment remain a considerable way off, largely because the hardware required to run quantum systems at scale is still enormously costly to build and maintain.

Disclosure: Sabrina Ortiz's travel to IBM Think was paid by IBM. The Deep View's coverage is editorially independent from the companies we cover.

TOGETHER WITH MIRO

Stop Wasting Sprints on the Wrong AI Features

If you've shipped a new feature, you know how it goes. Weeks of engineering work before anyone can actually see what you're building -- and by then, you're already headed in the wrong direction.

Prototyping fixes this. A fast, clickable version of your idea gets the whole team on the same page before anyone writes real code.

The catch? The prototyping tool landscape changes every few weeks, and most teams are just picking whatever feels right in the moment.

Miro put together a free guide to help with exactly that: a requirements checklist, a landscape overview, and a scoring framework so you can evaluate tools without the guesswork.

Download the free guide

PRODUCTS

ChatGPT accuracy improves in new model push

OpenAI's Instant model, the lightweight option available to everyday users, just got an upgrade.

On Tuesday, OpenAI launched GPT-5.5 Instant, which was updated to be more reliable, with the company claiming it boasts “significant improvements in factuality across the board.” This claim is particularly notable given that it is the default model for its hundreds of millions of daily users. 

Benchmark performance shows that GPT-5.5 Instant outperforms GPT-5.3 Instant, the model currently used in ChatGPT, across multimodal reasoning, document parsing, and science and math evaluations. In everyday performance, OpenAI says the updates mean that users experience: 

• Less hallucinations: GPT-5.5 Instant produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in areas like medicine, law, and finance and inaccurate claim reduction by 37.3% on challenging conversations users had flagged for factual errors

• More capable: The model is generally smarter and more capable across photo and image models, STEM questions, and choosing when to fetch information from the web 

• Tighter responses: More to the point responses that don’t lose substance

• More personalization: It is more effective at using context from either past chats or connected files and Gmail by being faster at searching past conversations. This is rolling out to Plus and Pro users on the web, and coming soon to mobile. It will expand to more plans in the coming weeks

OpenAI also introduced memory sources across all ChatGPT models, which give users the ability to see the context used to personalize their responses and then delete or correct it if something is outdated or not wanted to be cited. Essentially, it's giving users more control of how their past chats are used and referenced. 

The memory sources feature is rolling out to all consumer plans on the web and soon on mobile. GPT-5.5 Instant is rolling out to all ChatGPT users, replacing GPT-5.3 Instant as the default model, while paid users can still access GPT-5.3 Instant for three months before being retired. It is also available in the API as “chat-latest.”

As OpenAI mentioned in the release, because it is a model accessed by hundreds of millions of people every day, even the slightest upgrade is significant. That is specifically true when it comes to improving accuracy and reducing hallucinations. While it is extremely difficult to completely eradicate that issue, since AI models are not actually thinking but rather referencing past data to make predictions based on patterns, any reduction is valuable, as the risk of spreading inaccurate information is significant. 

LINKS

• Authors, publishing groups sue Meta, Zuckerberg over copyright 

• White House discusses possible government oversight on AI models 

• Apple considers using Intel, Samsung for main device chips

• Meta used bone structure analysis to find users under 13 on IG, Facebook

• Coinbase slashes 14% of its workforce, or 700 employees, citing AI

• Amazon CEO Andy Jassy said massive AI spending will bring major growth

• Webhooks in Gemini API: push notifications replace constant polling

• ChatGPT for Excel and Google Sheets: Add-on is now generally available

• ChatGPT images: Usage up >50% in just a few weeks since release of new model

• Claude Code update: 37 CLI changes, includes fixes across focus mode, MCP tools, etc

• OpenAI app: New iOS app for school and work organizations

• Perplexity Computer: The AI firm launched a computer use tool for professional finance

• Stripe: Forward Deployed AI Accelerator, Marketing

• CSL: Associate Director, AI & Advanced Analytics

• Beacon Hill Staffing Group: Sr Agentic AI/ NHI Security Engineer/ Architect – Remote

• TetraScience: Scientific Business Analyst, Scientific AI- Boston

GAMES

A QUICK POLL BEFORE YOU GO