Vladimir Troy

Just one month after announcing our partnership with OpenAI, we’re launching our first model together: OpenAI Codex-Spark, powered by Cerebras. Codex-Spark is built for real-time software development. In coding, responsiveness is the product. It is not a nice to have. Codex-Spark is optimized for targeted code edits, logic revisions, and frontend iteration. It gives developers near-instant feedback so they can stay in flow. Powered by the Cerebras Wafer-Scale Engine, it runs at over 1,000 tokens/s. That speed fundamentally changes the experience. We did not build this to win a benchmark. We built it so developers could move faster. I’m proud of how quickly the OpenAI and Cerebras teams have brought this to life. This is what fast execution looks like - deep engineering collaboration, rapid iteration, and shipping real products developers can use today. We are just getting started. When inference is fast, entirely new markets open up. We plan to lead that shift with our partners at OpenAI.

1,282

52 Comments

Our CEO Jensen Huang joined Y.L. Wang, TSMC vice president of operations, and Ray Chuang, CEO of TSMC Arizona, in Phoenix today to celebrate the first NVIDIA Blackwell wafer produced in the United States. This achievement marks a huge step forward in paving the way for sustained American leadership in AI. Onshoring world-class chipmaking to American soil is crucial to meeting the growing demand for AI. Within the next four years, NVIDIA plans to produce up to one-half-trillion dollars of AI infrastructure in the U.S. — collaborating with TSMC, Foxconn, Wistron, Amkor Technology, Inc., and SPIL 矽品. Manufacturing the world’s AI technology stack in America will fuel innovation and bolster the US supply chain. The production of American AI factories is projected to create hundreds of thousands of jobs and drive trillions of dollars in economic security over the coming decades. Read more: https://bit.ly/4oAvrzP

65

1 Comment

.@SambaNova steps up its challenge to Nvidia with new chip, $350M funding and a powerful ally in Intel https://bit.ly/4cbo8eX @SiliconANGLE @Mike_Wheatley “They care about performance. If SambaNova and Intel’s inference platform is competitive, the biggest challenge...” - @holgermu @constellationr

1

NVIDIA CEO Jensen Huang joined CSIS to discuss U.S. leadership in AI, including the importance of expanding energy infrastructure to support AI buildouts.

20

1 Comment

The Arm Total Design Ecosystem just keeps getting better🥳 Arm partner AMI has launched a new developer program for Arm Total Design. Powered by Arm Neoverse™ Compute Subsystems (CSS) this program offers a simple, cost-effective pathway for the Arm Total Design ecosystem to transform their unique designs into market-ready products. This is an exciting step for the Arm partner ecosystem on the journey to accelerating and streamlining Arm-based chiplet development. 💪⚡️ https://okt.to/PW6u3O

1

As AI and HPC reshape data center demands, deterministic, fabric-optimized connectivity is key. The connectivity bundled with your average server refresh is designed to connect components—not to maximize performance under pressure. If you specify ATTO, that changes everything. ATTO’s Gen 7 Celerity HBAs and FastFrame NICs support NVMe over Fabrics and intelligent Multipathing—enabling scalable, low-latency infrastructure that meets today’s performance needs and tomorrow’s data intensity. Our engineers know how to optimize connectivity for any workflow. Visit www.atto.com to learn more! And make sure you ask for ATTO when you spec your next server or infrastructure refresh. #DataCenterPerformance #AIInfrastructure #NVMeoverFabrics #HPCNetworking #connectivitymatters #perfromance #Ethernet #Fibrechannel #servers #Linux #vmware #Windows ATTO Technology, Inc.

26

3 Comments

𝗧𝗵𝗲 𝗹𝗮𝗿𝗴𝗲𝘀𝘁 𝗰𝗵𝗶𝗽 𝗲𝘃𝗲𝗿 𝗺𝗮𝗱𝗲 𝗶𝘀𝗻’𝘁 𝗳𝗿𝗼𝗺 𝗡𝗩𝗜𝗗𝗜𝗔 — 𝗮𝗻𝗱 𝗶𝘁’𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝗔𝗜 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲. A new episode of Startup Project is live, diving into one of the most consequential (and least understood) shifts happening in AI today: hardware architecture. In this episode, Nataraj sits down with Cerebras CEO & Co-Founder Andrew Feldman, widely regarded as the only independent chip startup competing head-on with NVIDIA and AMD. That positioning matters even more as the AI hardware market consolidates—most recently highlighted by NVIDIA’s acquisition of Groq. Why this conversation matters: → Cerebras built the largest chip ever manufactured using wafer-scale computing → Their system delivers up to 27× faster inference than NVIDIA-based setups for certain workloads → The company has signed a $10B+ compute deal with OpenAI, raised $1.1B at an $8.1B valuation, and is targeting a 2026 IPO Rather than scaling AI with ever-larger GPU clusters, Cerebras asks a more radical question: What if the GPU itself is the bottleneck? The episode explores: → Why GPUs start to struggle as AI inference scales → How wafer-scale computing changes system design → Why inference—not training—is becoming the real cost center → What it realistically takes for a startup to challenge NVIDIA and AMD 🎧 Listen to the full episode of Startup Project → YT: https://lnkd.in/gw3tzYc6 → Spotify: https://lnkd.in/gZ8cNmPd #StartupProject #AIInfrastructure #AIHardware #Semiconductors #Cerebras #NVIDIA #Inference #Startups

2

1 Comment

Silicon photonics learns to compute on quantum dots Source: Zhican Zhou and colleagues, "Integrated quantum dot lasers for parallelized photonic edge computing," Advanced Photonics 8(2), 026017 (2026), from Yating Wan's group at KAUST with John Bowers's lab at UCSB Modern AI burns electricity. Every ChatGPT query, every face-unlock, every self-driving frame is a mountain of multiply-and-add operations grinding through silicon transistors, each one bleeding heat. Optics has long promised a way out: photons don't push electrons around, they just travel and interfere, and you can pack many wavelengths down the same waveguide doing math in parallel. The catch is that nobody could build a clean, stable light source for this directly on a chip. So far the best optical "neural networks" have been laboratory rigs with a benchtop laser the size of a shoebox piping light into a tiny chip through fiber. The KAUST/UCSB team just put the whole stack on one 300 mm silicon wafer, the same kind your laptop processor is cut from. Their trick is a peculiar little device called a quantum-dot mode-locked laser. The active medium is a forest of nanoscopic InAs islands, each so small that electrons inside it behave more like atoms than like a bulk crystal. This gives the laser two superpowers. First, it shrugs off heat: it keeps lasing past 150 °C, where ordinary semiconductor lasers fall apart. Second, when "mode-locked," it doesn't emit one color but a precise comb of dozens of evenly spaced colors at once. Each color becomes a parallel computing channel. Why does that matter? In an optical neural net, "weights" are stored as tiny ring-shaped resonators that pluck out specific wavelengths. They're brilliantly compact but obsessively temperature-sensitive: warm the chip by a fraction of a degree and the rings drift off-resonance, and the network's accuracy collapses. In published demonstrations a two-degree thermal wobble can drag a digit-recognition task from 99% accuracy down to 67%. Most optical accelerators today spend a huge fraction of their power budget on tiny heaters fighting this drift. By moving the light source from a heat-isolated external box onto the same wafer as the rings, and using a laser that doesn't mind warmth in the first place, the team cuts optical losses by 2 to 6 dB (meaning a much fatter "photon budget" for the math), shrinks weight drift by more than half, and gets about 40% higher compute density than rival photonic chips. The bigger picture: this is the kind of work that decides whether photonic AI stays an academic curiosity or graduates into shipping products. Edge devices, phones, drones, cameras, satellites, AR glasses, aren't limited by raw FLOPS. They're limited by joules and heat. Light, if you can finally tame it on a chip, is the cleanest way out. KAUST and UCSB just showed the taming part might actually be a manufacturing problem now, not a physics one. https://lnkd.in/eDiFNvvs

37

1 Comment

Excited to announce our partnership with Alchip Technologies. Together with TSMC, we’re building an ecosystem to bring optical I/O and co-packaged optics into high-volume manufacturing — enabling scalable, power-efficient optical engines for hyperscaler AI scale-up networks.

101

2 Comments

Another great write up on the NVIDIA - Marvell Technology partnership. This time from Taryn Plumb at Network World. ▪️ Heterogeneity of AI silicon is inevitable as inference scales ▪️ But owning the connectivity that powers the AI environment is the next battleground ▪️ NVIDIA has clearly put its focus on NVLink enablement Check out Taryn's piece below - really interesting read ⬇️

11

Today we’re proud to announce our first commercial medium-scale AI data center in Santa Clara — the CSC1 site. This 35 MW facility is purpose-built for AI inferencing workloads. It will host multiple tenants and deliver a true state-of-the-art environment with capabilities that go well beyond standard data centers and traditional colocation. It is the first at scale demonstration of our Flex-Grid technology in action and production. From day one, CSC1 will support NVIDIA Vera Rubin systems and provide enterprise-grade high availability across power, cooling, networking, and operations. This is just the beginning. Over the next 24 months we plan to deploy multiple additional high-end inference data centers, further strengthening our leadership position and perfectly complementing our large-scale AI training portfolio in Texas. Space is still available. If you’re looking for reliable, high-performance data center capacity in 2026, 2027, and beyond, we’d love to talk. Reach out to our team at sales@edgecloudlink.com to learn more or secure your allocation today. https://lnkd.in/gFUqK8GS #AIDatacenters #AIInfrastructure #EdgeCloudLink #NVIDIAGPU #DataCenterExpansion #innovation #hydrogen #datacenter #greenpower #sustainability #missioncritical #informationtechnology #technology #tech #data #business #future #IT #networking #datacenter #thefutureisnow #hellotomorrow #leader #innovation #InternetEvolution #gigawatt #AI #change #future #edgecloud #edgecloudlink #nvidia #gb300 #VeraRubin #Inference ECL | Yuval Bachar | Michael Glickman | Hadar Feinberg | Colin Kincaid | Jacob Ben-David | Ron Gavriel | David Eduardo Cermeño Pinzon | Dean Nelson | Brian Hauge | Seshadri Sivakumar | Nathan Shuchami | Ben Omelchenko | Edith Hsu

213

33 Comments

Only 5 days until NVIDIA GTC! We're excited to join industry leaders in San Jose to explore how distributed infrastructure is reshaping AI deployment. ICYMI: Our CTO and Co-Founder, Sharon Zhang, will be speaking during the session "AI Grid in Action: Killer Applications for Distributed Inferencing That Telcos Can Monetize" on March 17. Alongside leaders from NVIDIA, Decart.ai, and Linker Vision, Sharon will discuss how distributed AI grids enable real-time, hyper-personalized applications and why inference closer to users is critical for latency, cost efficiency, and scalability. If you're attending GTC, we hope you'll join the session. Also, drop by Booth #3215 to say hi to the rest of the Personal AI team. 👋🏼 https://lnkd.in/eFhCdfJi

75

1 Comment

Cryptography algorithms are just a small subset of processor-intensive functions that benefit from hardware acceleration—but they’re among the most critical, given the pace of evolving threats. eFPGA lets you accelerate any algorithm, and unlike fixed ASIC hardware, it allows you to swap functions on the fly, enabling true run-time algorithm agility. Who doesn't want that? With our partners like Andes Technology, Codasip, Siemens and Rise Design Automation this technology is available now and they have expertise to enable your ASIC or SoC with this capability. Contact me to learn more!

5

The most impactful and important SDK release Trainiun Neuron team has worked on since inception: Native PyTorch support for Trainium, open source kernel (NKI) compiler and pre-optimized NKI kernel library, Neuron Explorer (IMO, the fastest and most intuitive ML profiler in industry), optimized inference implementation for major MoE and video models (GPT-OSS, Qwen, Pixtral….), complete vLLM v1 support !! All with full support for Trainium3. Using Trainium, whether for research, training , or large scale inference just became more intuitive and simple. https://lnkd.in/gjfhjv4j

626

10 Comments

January 2026 delivered a quiet but consequential infrastructure shift. AMD’s ROCm stack is now a first-class platform inside the vLLM ecosystem, a change that matters far beyond benchmarks or branding. vLLM began at UC Berkeley as an effort to make large language model serving cheaper and more resilient, built on the idea that inference should not be locked to a single vendor. That philosophy has since scaled into a global open-source system running across hundreds of thousands of GPUs, governed by the PyTorch Foundation and shaped by a community that values hardware pluralism over habit. Earning first-class status required execution, not announcements. Over eight weeks, AMD engineers pushed ROCm support from fragile to production-ready, with CI passing more than 90 percent of tests, official Docker images live, pip-installable wheels available, and day-zero support in vLLM Omni. Quantization kernels, KV cache performance, and multimodal paths were upstreamed rather than promised. Character.AI doubled inference throughput on MI300 hardware, while Red Hat and DigitalOcean built enterprise and cloud support around the stack. The significance is structural. CUDA gravity remains real, but it is no longer unchallenged by default. With vLLM reaching this level of operational maturity on AMD hardware, inference infrastructure moves from vendor loyalty to economics and execution. For Tech News watchers, this moment signals a shift from theoretical competition to practical choice, where hardware decisions become pricing conversations instead of ideological ones. #AIInfrastructure #OpenSource #Inference #Compute #CloudInfrastructure

Analysts are talking about AI-first agencies for a reason. AI is not just improving productivity, it is changing how outcomes get delivered in services. Thiago Maior has been operating in DevOps for 20+ years and has a clear point of view on what it looks like when AI runs ops and engineers steer strategy. If you are building a services firm, this is one worth hearing. Conference website link in the post.

13

1 Comment