Marc Fleischmann

NVIDIA Unveils Rubin Generation Platform at CES 2026: Redefining Data Center Efficiency with 45°C Liquid Cooling At CES 2026, NVIDIA CEO Jensen Huang introduced the Rubin generation data center platform, marking a significant evolution in AI infrastructure design. Building upon the Blackwell architecture's liquid cooling foundation, Rubin implements an advanced 45°C warm water single-phase direct liquid cooling system that fundamentally reimagines thermal management in high-performance computing. Key Technical Achievements: 🔹 2x power density compared to Grace Blackwell 🔹 5x peak inference performance improvement 🔹 3.5x peak training performance enhancement 🔹 1.7x increase in transistor count 🔹 Consistent airflow and water temperature requirements As Huang emphasized: "This is a miracle. The airflow coming in is roughly the same. More importantly, the water temperature coming in is the same at 45°C. At 45°C, no chiller is needed to cool the data center. We're essentially cooling this supercomputer with hot water, which is incredibly efficient." Industry Implications: The Rubin platform addresses three critical challenges facing modern AI infrastructure: 1. Energy Efficiency: By eliminating the need for chillers, operational energy consumption is dramatically reduced 2. Thermal Management: Higher flow rates and optimized cooling prevent thermal throttling, ensuring consistent peak performance 3. Sustainability: The system's ability to operate with warm water significantly reduces environmental impact This innovation arrives at a crucial juncture as AI workloads continue to scale exponentially. The ability to double power density while maintaining thermal efficiency positions Rubin as a cornerstone technology for next-generation AI factories. The strategic shift toward warm water cooling represents not just an incremental improvement, but a paradigm change in data center design philosophy. As organizations invest billions in AI infrastructure, technologies like Rubin will be essential for achieving both performance and sustainability objectives. What are your thoughts on the future of liquid cooling in enterprise data centers? How will this impact your organization's AI infrastructure planning? #NVIDIA #CES2026 #Rubin #DataCenter #AI #LiquidCooling #EnterpriseAI #Sustainability #HighPerformanceComputing #AIInfrastructure #JensenHuang #TechInnovation

2

More insights coming out of Blue Yonder ICON from Nucleus Research. Chris Burchett, Matt Ellsworth, Lindsey Hiefield, Sarah Hart, Rebecca Livingston

8

AI infrastructure is evolving quickly, and NVIDIA is making one thing clear: shared KV cache is becoming essential for the future of inference. But the real challenge is figuring out how to get there from today’s GPU environments without slowing innovation or starting from scratch. Betsy Chernoff shares how WEKA helps teams take a practical path forward, delivering performance gains now while building toward AI Factory and ICMS architectures. If you’re thinking about where AI infrastructure is headed, this is worth a read: http://spr.ly/6046h7rHn

5

Lithography the process of putting an image onto the surface of the silicon. Pretty much like the way a silkscreen puts  “Team Building Exercise 1999” on a T-shirt. Except that this image has to be the highest resolution, with the smallest microscopic features, of anything humans produce. “Moore’s Law” usually refers to increasing transistor density. Basically, how can we make transistors half as big as they were 18 months ago? Every time we figure that out, computers get twice as powerful. The state of the art uses Extreme Ultraviolet (EUV) light to do the lithography. The machine that can do this cost $50 billion to develop. It has 500,000 parts. Only the Large Hadron Collider is more complicated. To buy one costs $250 million and you’ll be stuck on a waiting list that is $40 billion long. The machine comes from ASML in the Netherlands and they don’t have a single competitor, in the entire world. That machine shoots a tiny ball of molten tin into a vacuum and blasts it with two lasers. This produces a flash of 13.5 nanometer ultraviolet light that gets aimed at the surface of a silicon wafer. ASML advanced from 193nm to 13.5nm light to make this possible, but there’s a problem. The diffraction limit of 13.5 nanometer light was set by either God or Issac Newton and there’s nothing we can do about it. We can’t print features smaller than that and there’s no practical way to do lithography with a shorter wavelength. When people say that Moore’s Law is over, this is why. The semiconductor industry knows this, so they’ve tried to solve the problem by handing it over to the marketing department where the laws of physics don’t apply. You’ve seen them progress from 45nm to 30nm to 20nm over the last decade, then all of the sudden, 12nm, 7nm, 5nm & soon 3nm chips are coming. This is all just marketing bullshit. This measurement in chips used to be half the distance between the centers of two features. Once marketing took over, they started measuring half the distance between the edges of two features. Instant improvement! Then they started measuring other random stuff. Other kinds of improvements in chip design helped to gloss over the fact that we are no longer able to shrink the size of transistors. Today, there are extraordinary geopolitical machinations to control chip production. The U.S. has tariffs and export controls akin to those for fighter jets and ICBMs. Access to chip production is as critical to superpowers as oil. Lace Lithography had been in stealth since we invested in them a few years ago. They’ve invented the technology that can go beyond Extreme UV and put Moore’s Law back on track. By using helium atoms instead of light, they can make transistors 10x smaller than the physical limit of ultraviolet light can. https://lnkd.in/dq8ruxHV

100

6 Comments

The potential for AI in 3DIC design is staggering, but there are challenges that must be solved. I’ve been working with our teams implementing the upcoming AI-driven flow for the Cadence 3DIC platform. While the gains in productivity, test bench coverage and code quality improvement are undeniable, three massive hurdles remain for the industry: The Cost of Intelligence: Sophisticated multi-agent workflows (like Claude 4.6) are powerful but expensive. A burn rate of $10k–$20k/month on tokens is easily reached with moderate usage. The Latency Gap: Commercial model latency is still a bottleneck. Between GPU deployment waits and the need for frequent context compaction, the flow isn't always fluid. The Security Vault: In the world of high-end silicon design, IP is everything. Cloud-based LLMs still raise major red flags for customers regarding security and data protection. The "winner" in the EDA AI race won't be the one with the best model, it will be whoever solves the Cost, Speed, and Security trifecta. #Cadence #EDA #3DIC #Semiconductors #ChipDesign #AdvancedPackaging #AI #GenerativeAI #HardwareEngineering #Innovation

57

https://lnkd.in/eiuhkTb8 Rüdiger Dornbusch, a German-born economist known for his work on exchange rates and international macroeconomics. He famously said: “In economics, things take longer to happen than you think they will, and then they happen faster than you thought they could.” This insight captures the paradox of technological and economic change: long periods of buildup followed by rapid, often disruptive acceleration. It’s especially relevant in contexts like AI, where foundational research may simmer for decades before breakthroughs suddenly reshape industries. In 2012, a group of researchers at Google Brain, led by Andrew Ng trained a deep learning neural network to identify cats by looking at over 10M YouTube thumbnails. At the time, anyone paying attention thought this would change the world. Most thought it would happen within 5 years. 10 years later, ChatGPT 3.5 launched, unleashing the democratization of innovation with AI for the first time. Even still, we are coming up on 3 years beyond the event and enterprises are still struggling to drive adoption and, in some cases - even identifying and prioritizing compelling use cases. While this announcement (with more to come) may sound like we are finally approaching the singularity, please note it will take years from now to launch this 10 GW project. David Linthicum Bruce Coughlin John Treadway Troy Angrignon

18

1 Comment

Nvidia just delivered another blow-out quarter. $57B in revenue (+62% YoY), $51B from data center alone, Blackwell chips sold out for the next twelve months, and guidance lifted to $65B for Q4, and CEO Jensen Huang spoke about a visible $500B+ order backlog stretching into 2026. The fundamentals have never looked stronger. Yet the stock barely held its after-hours pop, and the entire AI trade was hammered the very next day. Why? Because the narrative machine is now louder than the numbers. “AI bubble” rocketed to the top tail risk in Bank of America’s latest fund-manager survey, every 2-3% dip instantly becomes headline clickbait, and social media turns it into a self-fulfilling stampede before anyone even reads the earnings transcript. Moves from big money only pour fuel on the fire. SoftBank’s Masayoshi Son, the poster child of AI optimism, sold his entire $5.8B Nvidia position last month to recycle capital into other ventures. While Peter Thiel's hedge fund also recently sold its entire $100 million stake in AI chipmaker Nvidia. And while everyone was screaming “bubble burst,” Warren Buffett’s Berkshire Hathaway silently initiated a brand-new $4.3B stake in Google during Q3, a move that barely made a ripple in the press. So in one corner we have sold-out factories and half-trillion-dollar backlogs; in the other we have headline-driven volatility, slowing hyperscaler capex growth, and legendary investors either cashing out or choosing the company that’s actually monetizing AI at scale right now. Same data set, completely opposite conclusions. What nobody can seriously dispute right now is that every single GPU Nvidia makes is already sold before it even leaves the factory. The vast majority of companies are still just dipping their toes in with small AI pilots, and hardly anyone is using it at full scale yet. Power shortages and chip supply bottlenecks are baked in until at least 2027. And crazy as it sounds, a headline screaming “AI bubble” can now wipe billions off stock prices faster than the strongest earnings report can add them back. Forget picking the exact high or low. Everyone’s really just waiting to see which side comes out on top: the sold-out factories and order books, or the panic headlines. What’s the one indicator you’re tracking right now to decide if this is the next internet or the next dot-com mania? Curious where everyone stands this week. Disclaimer: This is informational only, not financial advice. Investing carries risks; consult an advisor. #nvidia #earnings #wallstreet #investment #stockmarket #ai #bubble #tech #gpu #smartmoney

4

If intelligence is the log of compute… it starts with a lot of compute! And that’s why we’re scaling our GPU fleet faster than anyone else. Just last year, we added over 2 gigawatts of new capacity – roughly the output of 2 nuclear power plants. And today we’re going further, announcing the world's most powerful AI datacenter, located in southeastern Wisconsin. Fairwater is a seamless cluster of hundreds of thousands of NVIDIA GB200s, connected by enough fiber to circle the Earth 4.5 times. It will deliver 10x the performance of the world’s fastest supercomputer today, enabling AI training and inference workloads at a level never before seen. For AI training workloads, you need compute at exponential scale. That’s why we designed the datacenter, GPU fleet, and network together as one integrated system. This ensures a single job can run from day 1 at exponential scale across thousands of GPUs.  Fairwater uses a liquid-cooled closed-loop system for cooling GPUs that requires zero water for operations after construction. And we’re matching all of the energy that is consumed with renewable sources. And of course, it is just one of several similar sites we’re lighting up across our 70+ regions. We have multiple identical Fairwater datacenters under construction in other locations across the US, in addition to our AI infrastructure already deployed in over 100 datacenters around the world, powering model training, test-time compute, RL tuning, and real-time inference at global scale. Too often during times like this, people go with the current and only later wonder, how did we get here?  With Fairwater, we're charting a new path: doing the hard engineering work, bringing compute, network, and storage into one highly scaled cluster, and designing closed-loop energy systems to meet real-world computing needs. And partnering with local communities to ensure it's thoughtfully done in a way that is sustainable, creates new jobs, and expands opportunity.  We are thrilled to see this take hold in Wisconsin, and we are just getting started. Learn more about Fairwater here: https://lnkd.in/gpdni9gt

7,648

386 Comments

“What does NVidia’s effective acquisition tell us about the semiconductor industry? Data movement is the problem !” There is a lot to like about Groq’s LPU, but the most interesting thing in my opinion is Groq’s choice to use SRAM instead of DRAM. Why ? It turns out that the limiting factors - both in energy and speed - is moving data from CMOS to HBM. A majority of heat is generated through this data transfer. Even the SRAM is way bigger than DRAM and more expensive, it appears to be worth it because of the higher throughput and lower power consumption. It also turns out that packaging with HBM is very complex too and very difficult to test. So by using SRAM only, there is a good chance that we will see significant packaging reduction costs, even though SRAM is much expensive. Another note - HBM GPUs are still better for training (simply because there is lots of memory), but because adding more parameters to LLMs does not necessarily improve performance, its not really that necessary. What’s really going to make the difference in training in the future is synthetic data generation, which LPUs with SRAM appears to also have the advantage. My prediction is that future training clusters will have large percentage of compute move from doing SGD to synthetic data generation. I’m quite interested in the dataflow SDKs that LPUs are building too. (Note Cerebras has a similar architecture). I think this is going to be a major theme in 2026 and beyond. What does that mean ? I’m long on TSMC and not so bullish on Samsung and Hynix. Also, I think there is going to be a huge push in the future to develop RRAM / MRAM alternative to SRAM. This is going to be VERY interesting. If anyone know any MRAM groups that are building AI accelerators, please put me in touch ! Groq NVIDIA #LPU #GPU

25

2 Comments

🎙️ “AI isn’t a product — it’s a velocity enabler.” — Matt Siemens, NUSO At the #Crexendo User Group Meeting, I caught up with Matt Siemens of NUSO, where the company is taking a bold, practical approach to AI and expanding across Europe. 🌍 In his main-stage talk, Matt explained how NUSO has built a digital twin of its own network — a live, structured data model that keeps AI grounded in real-world conditions. The result: 🤖 Fewer hallucinations 🧠 Smarter transcription and sentiment analysis 🔁 Better context for every customer interaction Matt also highlighted how #vCons** act as “micro-mirrors”* — threading together the moments that define customer experience and helping AI focus on outcomes, not just outputs. Beyond AI, NUSO announced a global technology partnership with Crexendo** and continued European expansion — now with operator authority in 14 countries, and new momentum in 🇫🇷 France, 🇩🇪 Germany, 🇪🇸 Spain, and 🇮🇹 Italy. If you’re at **Crexendo UGM**, stop by the **NUSO Lounge on the 3rd floor** for a quiet conversation and a closer look at their next-generation UCaaS platform. Or learn more at 👉 [nuso.cloud](https://nuso.cloud) #NUSO #AI #DigitalTwin #vCon #UCaaS #TelecomInnovation #CrexendoUGM #TechnologyResellerNews #CloudCommunications #AIinTelecom #CrexendoUGM 2025 Snom Technology GmbH\ #SnomAtTheShow Doug Green TR Publications #Podcast Julie Smith Chad Collins Antoine Karachekhlian Vincent Gianfrancesco Outsourcing Services International | OSI TELCLOUD NUSO Yealink Nomadix, an ASSA ABLOY company Grandstream Networks Ozonetel | oneCXi 888VoIP Poly HP NTSDirect VOLA NETWORKS FaxSIPit Services Inc. CCT Solutions Adaptiv Networks Creo Solutions Ecosmob Technologies Sinch TransNexus Amanda Howard Jason Byrne

26

2 Comments