Raymond Chik

Today's Artificial Analysis data on OpenAI GPT-OSS-120B inference shows Cerebras Systems is fastest whether for one request or 10 requests. The data does not lie: wafer-scale AI compute is unmatched when it comes to large-model inference. No number of GPUs can play here. Source in comments.

297

5 Comments

How much of a Headstart will Meta have in the Display AI Glasses Race? In a few weeks, #Meta is launching their #Hypernova Glasses - the first major brand to sell #AIGlasses with a display, giving them first mover advantage in the race to establish the next major computing platform. https://lnkd.in/gxgf4d8g How much of a lead will they have before they face competition from other major brands? #Samsung is working on Project Haean.  Final specs have not been published yet, reports are it will be available late 2025. These glasses will run #AndroidXR, which in partnership with Google and #Qualcomm,  https://lnkd.in/dXqMDFwQ #Google has demonstrated a prototype running their Astra AI assistant.  @ShahramIzadi has hinted at some of its features in his blog post, and, best guess is that this will launch in the second half of 2026. https://lnkd.in/dvmV-qyz #Snap is planning to release a consumer level product called ‘#Specs’ according to @EvanSpeigel’s keynote address at AWE.  Details are thin, and some doubts have been raised about Snap’s ability to internally fund this project, but the release date is set for late 2026. https://lnkd.in/d8i-B5SS #Apple is working on project N50.  Shrouded in Apple secrecy, the project is reported to have more than one glasses embodiment.  @MarkGurman, has reported that they are developing a custom IC with mass production slated for late 2026 or early 2027.  This would put Apple Glasses hitting the market about two years from now. https://lnkd.in/dF5zdNwU #Amazon is working on a project called ‘Amelia’ specifically tailored towards their own use case.  It is not clear that these glasses will be sold commercially, but they are potentially a formidable competitor. https://lnkd.in/dvnCzHFW #Microsoft, who has exited the #AR hardware space may still be in the game.  According to The Verge, they have a deal with Samsung to procure components , coming to fruition in 2026.  This has to be seen as speculative, given Microsoft’s strong step away from wearable hardware.  https://lnkd.in/dXbYKUNu Absent from this list are giant companies from China that are rapidly developing devices.  #Alibaba has already announced a product, and there are rumors that #ByteDance, #Baidu, and #Xiaomi are developing their own models.  Follow @DovidShick to stay updated with the latest AI smart glasses news.  Subscribe to my newsletter for more in depth stories and analysis: https://lnkd.in/dJhJbExV 

114

27 Comments

The Next Leap in AI Networking: Co-Packaged Optics As AI factories scale, NVIDIA Photonics is redefining data center networking by: ✅ Removing the need for pluggable transceivers ✅ Delivering more resilient and energy-efficient networks ✅ Enabling co-packaged optics for both Quantum InfiniBand and Spectrum-X Ethernet The result? Faster time to first token and scalable, sustainable infrastructure to power the era of agentic AI. 👉 Watch the NVIDIA Quantum-X Photonics switch come to life in an #AIFactory: https://bit.ly/48gtPWW

94

2 Comments

Turning traditional data centers into artificial intelligence (AI) factories. 🌐   In a recent blog from NVIDIA, they highlighted the NVIDIA Omniverse Blueprint for delivering high-performance, energy-efficient infrastructure optimized for digital success. #OurJacobs is serving as the design integrator, incorporating the physical and digital layers for seamless, resilient and long-term results.   While buildings used to be designed independently of the platforms within them, learn how we’re working together to think differently at the link: https://lnkd.in/gRhF2Ukc   #AdvancedManufacturing #DataCenters #ChallengeAccepted #DigitalTools

252

2 Comments

The Phaidra team could not be more excited and humbled to partner with NVIDIA on the Reference Design for GW-scale AI Factories. When I joined Joseph Kava's Google DC team back in 2010 as a fresh college grad, the industry was still talking about shipping containers and "chicken coop" designs as the new thing. Google had begun building 30MW modular DCs, which many folks at the time thought was crazy. Who could possibly need 30MW of compute?! How the world has evolved. Instead of building 5MW shipping containers or 30MW DCs, we're now building 1GW+ AI Factories. Purpose-built infrastructure designed from soup-to-nuts for AI workloads. And as the next generation of AI infrastructure becomes bigger, more dynamic, and more complex — we're going to need equally ambitious leaps in technology to control and optimize these AI factories. Like AI agents to orchestrate the cooling, power, and workload management systems as a single integrated unit of compute. Thank you Vladimir Troy Andria Zou Wade Vinson Scott Wallace Seth Lachica (and many others!) for the deep partnership and collaboration. Now let's build! https://lnkd.in/gRhF2Ukc

133

5 Comments

I wrote a report on the GPU market, covering rendering, gaming, and HPC industries. Some interesting takeaways: - 3+ billion active gamers worldwide - 1+ billion GPUs shipped every year (the majority being smartphones) - 66% of the discrete GPU market has ASPs over $970 And here's the big one: - 25-50% of time and money is spent waiting for GPUs to render or perform simulations - *Over $1B/yr is spent waiting for your GPU to finish its work* - Faster GPUs save time/money - We only measured this $1B/yr number for content creation industries! Read here: https://lnkd.in/gdPrn_d9

32

5 Comments