AMD doubles down on design as Intel reels

Welcome to Runtime! Why AMD's $5 billion bet on data-center systems design could make Intel's problems worse, OpenAI brings fine-tuning to its most powerful enterprise model, and the latest funding rounds in enterprise tech.

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

Systems thinking

During the first ten years of the cloud computing buildout, Intel was pretty much the only chip supplier that mattered to data-center customers. As AI workloads have started to change data-center requirements over the last five years, its rivals have pounced.

Nvidia's ascent to the top of the mountain is well documented at this point, but AMD's data center division has also enjoyed enormous success in recent years thanks to being in the right place at the right time as AI demand surged and Intel stumbled. This week it unveiled plans to keep that momentum going with a $4.9 billion acquisition deal for ZT Systems.

• ZT Systems — "probably the biggest server maker you have never heard of," as The Next Platform put it — designs and manufactures servers, racks, and clusters for some of the biggest cloud computing providers in the world.
• AMD plans to spin off the manufacturing part of its business some time next year after the deal closes rather than compete with Dell and Supermicro, but it will keep the industrial design teams.
• AMD CEO Lisa Su told Reuters that it will use those designers to boost GPU sales by putting them to work on improving the company's ability to sell data-center customers a complete package of the components they need to run AI workloads at scale.

Server hardware designs were relatively simple and cookie-cutter during Intel's heyday, when it commanded more than 90 percent of the market for data-center CPUs. But as companies have started to depend more on AI workloads, those designs have become more complicated now that other chips now play an equally important role in overall system performance.

• Nvidia realized this years ago as it saw a steady increase in demand for its GPUs, which until that point were used to run video games, among enterprise data-center customers.
• But while data-center operators welcomed the performance boost, putting together reliable clusters that balanced the needs of CPUs and GPUs, which process information quite differently, was challenging.
• Nvidia launched its first DGX server designs in 2016 and continued to upgrade them as high–performance networking chips started to become the third leg of the data-center stool.
• The biggest cloud providers tend to build their own servers to their own exact specifications, but chip makers understand how their products fit together at a deeper level, and AMD wants to make sure that potential customers understand how to incorporate its chips as part of their data centers.

Nvidia was foremost on AMD's mind when it made the ZT Systems deal, given the strategic importance of GPUs to data-center customers at the present moment. But it's also a blow to Intel, which really needs to come up with something new to regain its momentum in enterprise computing.

• Intel's current problems started with manufacturing mishaps, but it also hasn't designed a chip that forced the industry to look its way in a long time.
• Christopher Kelly, who retired from Intel last December after nearly 30 years at the company, wrote on LinkedIn last week that "rivals' products have caught up and in some cases surpassed Intel products in pretty much all product segments" and that the company has suffered an "alarming level of talent loss."
• ZT Systems could have brought fresh design thinking into Intel at a time when it desperately needs something to regain momentum.
• Instead, it will continue to watch its longtime rivals define the future of data-center chips, a market it once had pretty much to itself.

When do we get autotune

The powerful large-language models that upended every enterprise tech product roadmap in late 2022 actually require a fair amount of customization to make them useful inside most companies. This process is known as fine-tuning, and as of Tuesday OpenAI's corporate customers can start putting its latest model to work on their data.

Calling it "one of the most requested features from developers," OpenAI's GPT-4o now supports fine-tuning, the company announced in a blog post. "Fine-tuning enables the model to customize structure and tone of responses, or to follow complex domain-specific instructions," OpenAI said, and those custom tweaks could make LLMs far more useful for business applications frustrated by clunky one-size-fits-all models.

“We’ve been extremely focused on lowering the bar, the friction, the amount of work it takes to get started,” Olivier Godement, the head of product for OpenAI’s API, told Bloomberg. The announcement creates another interesting "frenemies" dynamic with OpenAI investor Microsoft, which announced a preview of fine-tuning for GPT-4 at Build in May.

Enterprise funding

Eppo raised $28 million in Series A funding to expand its cloud service, which allows customers to run A/B tests on new features.

CodeRabbit landed $16 million in Series A funding for its AI-powered code-review service.

The Runtime roundup

Palo Alto Networks beat Wall Street estimates and raised its fiscal-year guidance for both revenue and profit, citing increased demand for cybersecurity software.

CISA warned that threat actors are targeting a previously disclosed vulnerability in Jenkins, the widely used CI/CD tool, that users have failed to patch.

Microchip Technology reported that a cyberattack had disrupted its operations and delayed the production of low-level chips used in everything from cars to weapons systems.

Cisco is forcing employees to wait a month before learning if they'll be laid off in its latest round of cuts, according to TechCrunch, which should make for a super-productive month at Cisco.

Thanks for reading — see you Thursday!