Six Five Media

Why does OpenAI need to design its own chips? For Richard Ho, the answer is control. At Global Semiconductor Alliance, Patrick Moorhead caught up with the OpenAI silicon leader behind the recently announced Jalapeño chip to talk about why OpenAI is making the long-term bet on custom silicon. Ho says this isn’t an “or” strategy. It’s an “and.” General-purpose accelerators will remain critical, but designing silicon specifically for large language models gives OpenAI greater control over the infrastructure underlying its models and capabilities. And the bet extends beyond hardware. Asked what people are getting wrong about AI, Ho pointed to the tendency to jump from rapidly improving capabilities to the most extreme outcomes. His view is more grounded: look at what AI can actually do today, how people are using it, and how those capabilities are evolving. Inside OpenAI, he says AI is already helping engineers do more, creating “superpowered engineers,” rather than simply replacing them. That leads to a much bigger possibility: AI not just as another industrial revolution, but as the beginning of a new renaissance in human creativity and capability.

From grid hogs to grid assets. That's how Christopher Wellise, VP of Sustainability at Equinix, describes the shift underway at its data centers. Equinix now runs more than 100 megawatts of Bloom Energy fuel cells across over 19 sites in six states, with some locations aiming for 100% on-site power. The fuel cells can also move to biogas or hydrogen as those supplies grow. His bottom line for the AI era: sustainability has to be a core design constraint, never a side program. Watch the full session with Nick Patience: https://lnkd.in/gkaRxHmR #Sustainability #DataCenter #AIInfrastructure #FuelCells #SixFiveSummit

How do you design a chip for years from now when AI models are changing every six months? That’s one of the big questions Patrick Moorhead is digging into at the Global Semiconductor Alliance (GSA) event, where leaders from across the semiconductor industry are gathering to discuss what comes next. Patrick will moderate a panel featuring Paul Cho of Samsung Semiconductor, Richard Ho of OpenAI, Charlie Kawwas of Broadcom, and Mark Papermaster of AMD on one of the industry’s toughest challenges: How do you lock in a design, architecture, specification, or packaging strategy years before a product ships when AI models, ecosystems, supply chains, and infrastructure requirements are evolving so quickly? Stay tuned for Patrick’s coverage and insights from GSA.

AI data centers need more than faster chips. Power, cooling, networking, and hardware all have to scale together. At Open Compute Project Foundation Global Summit 2026, the open compute community will explore the designs and standards shaping the next generation of AI infrastructure. Six Five Media analysts Brendan Burke will be on-site, connecting the announcements to the practical decisions facing organizations building and operating AI data centers. Follow Six Five Media for coverage from the summit. https://lnkd.in/d7E_NNQz

Has Meta finally cracked mainstream AI wearables, or has it built a great agent that people mostly use on their phones? Patrick Moorhead and Daniel Newman take on that question this week on The Flip after Meta Connect 2026. Pat argues Meta has cracked the formula without needing full AR. Wearables have spent 14 years misfiring, from Google Glass to early AI gadgets that were slow and inaccurate. The new Ray-Ban Meta Audio glasses look like ordinary frames, and pairing them with the Muse agent delivers real utility through audio alone. He also sees promise in the Muse Charm, a keychain-sized device that carries Muse without glasses. Dan's counters: "Meta did crack the agent. It just didn't crack the wearable." The demand signal for Muse shows up in app store downloads on phones, most people wear AI glasses to listen to music, and the Charm hasn't shipped yet. When agents start paying mortgages or booking vacations, he expects people to reach for a screen and review the commitment first. For device makers and platform builders, the debate turns on whether AI changes consumer habits or rides existing ones. Dan's bar is whether people start leaving their phones at home. #AIWearables #SmartGlasses #AgenticAI #ConsumerTech #Meta #SixFivePod The Flip Disclaimer: The Flip is a simulated debate segment in which Patrick Moorhead and Daniel Newman argue opposing sides of a topic. The positions presented are for discussion and entertainment purposes and may not reflect the hosts' personal views.

Google is now processing more than 3 quadrillion tokens a month, a one followed by 15 zeros, and that volume is 7x higher than a year ago. Daniel Newman suggested that on the current curve, the figure could reach 3 quadrillion a week within just 1 year. Mark Lohmeyer, VP and GM of AI and Computing Infrastructure at Google Cloud Security, agreed that it probably will. Speaking in the AI Infrastructure track at the Six Five Summit 2026, Lohmeyer brought the number back to the enterprise. Most organizations grow consumption at a 7x rate, but they will still need to scale applications and infrastructure quickly while industry-wide demand for AI compute outpaces supply. That puts platform selection at the center of AI planning. Lohmeyer's advice is to think deeply about partners and platforms, and to choose ones that can scale as business needs grow and change. Watch the full session: https://lnkd.in/efjFf8Yg #AIInfrastructure #GoogleCloud #EnterpriseAI #AIInference #SixFiveSummit

Samsung Semiconductor President Paul Cho believes AI’s next bottleneck depends on how efficiently we use memory. At GSA, Patrick Moorhead asks what the AI industry needs to get right today to support the massive growth expected through 2031. Cho's answer goes beyond simply adding more memory. As AI systems become increasingly heterogeneous, the industry will need deeper memory hierarchies spanning HBM, LPDDR, and other technologies, along with software capable of making the best use of those resources. That also means tighter co-design across memory, GPUs and XPUs, networking, software, and emerging architectures rather than optimizing each layer in isolation. 2031 may look very different from today, but the architectural decisions that enable it are being made now.

For a while, the AI silicon market was judged by a simple scorecard: win one of a handful of major custom XPU design slots, or you’ve lost. Marvell Technology Chairman and CEO Matt Murphy never saw the opportunity that narrowly. Beyond the headline XPU wins, Marvell has built a portfolio of custom silicon and “XPU attach” wins spanning critical functions around the accelerator, from networking and memory expansion to security. Murphy says those opportunities now number in the dozens. Add Marvell’s connectivity and switching portfolio, and its strategy extends across even more of the AI infrastructure stack. His read on the market today: “This market is not a zero-sum game.” Watch the full session with Daniel Newman and Patrick Moorhead: https://lnkd.in/giMRFVZw

Compute workflows cycle, but data compounds. Tim Rausch, SVP of Product Engineering at WD, explains why that distinction should reshape how enterprises scale AI. Every AI interaction runs on a GPU and produces something, whether it's an email, a presentation, or a photo, and that output eventually lands on a hard drive. The next user runs on the same GPU resources and adds their data on top, and then the next user does the same. Rausch tells Matthew Kimball that compute gets reused while the data keeps compounding, which is why storage has to grow alongside GPUs. For infrastructure leaders planning AI buildouts, capacity planning belongs next to compute planning, because the data footprint keeps expanding long after the compute cycle ends. Watch the full conversation: https://lnkd.in/gaAu2YyK #AIInfrastructure #DataStorage #HDD #DataCenter #AIInfraSummit

Physical AI gets proven by deployed units, and most of what gets demoed never reaches the field. Muneyb Minhazuddin, Customer Growth Officer at Ambarella Inc., tells Brendan Burke and Matthew Kimball that buyers should measure what it takes to roll out tens of thousands of devices that run reliably for years, take upgrades, and keep getting smarter. Ambarella has shipped more than 50 million AI chips into cameras, cars, drones, robots, and wearables, well before the category had a name. Scaling to millions of devices calls for different engineering than scaling to a few hundred data centers. Minhazuddin points to latency. A 30 to 40 millisecond delay on a web checkout goes unnoticed, but a high-speed train braking for an obstacle needs a response in microseconds. He argues that physical AI therefore has to be re-engineered to run deterministically, the same way, millions of times over, instead of squeezing cloud models onto smaller hardware. For industrial buyers, the evaluation moves from demo quality to fleet-scale operational readiness: performance per watt, performance per dollar, and deterministic behavior under real-world conditions. Watch the full conversation: https://lnkd.in/gjx7VTUH #PhysicalAI #EdgeAI #Semiconductors #AIInfrastructure #AIInfraSummit