Marc Fleischmann

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/6045hYo1U

10

Every partnership starts with optimism. The new deal between OpenAI and Broadcom is no different. OpenAI needs more power to train bigger models. Broadcom knows how to build the hardware that makes it possible. Together, they’re betting on custom chips to keep pace with AI’s growing appetite for compute. But here’s the truth, most partnerships don’t fail in the lab. They fail in the boardroom. When deadlines slip. When priorities change. When one side starts protecting its turf instead of protecting the goal. If OpenAI and Broadcom can stay aligned through those moments, they’ll set the new standard for how tech alliances work. If not, this deal becomes another headline that aged too fast. If you’ve enjoyed this look at how partnerships really work, follow for more perspectives on the deals shaping the future of business and technology. #StrategicPartnerships #OpenAI #Broadcom #AI #Leadership #BusinessStrategy #Innovation #Partnerships #TechIndustry #Growth

17

1 Comment

Excellent piece in The Economist on Starcloud and orbital data centers today! 👌🤓 "Starcloud, a company founded in 2024 to pursue the idea of orbital data centres, has been crunching the numbers for AI in orbit—literally. In November the company sent Starcloud-1, a fridge-sized satellite containing an ordinary Nvidia H100 GPU, of the type used in AI data centres, into space. Starcloud used it to train a small AI language model, NanoGPT, on the works of Shakespeare, and to answer some queries while running Gemma, an open-source large language model made by Google. That provided valuable data on the reliability of AI chips under orbital conditions. The company also has a good handle on the other crucial figures. Start with specific power. Mr Andrew McCalip’s figure of 37W/kg comes from the thousands of satellites used in SpaceX’s Starlink constellation, thought to be state of the art, which provide high-speed internet around the world. But Starlink satellites have to do things that AI satellites do not. They need costly “phased-array” antennas to communicate with the ground, and must have a stable orientation at all times. AI satellites, by contrast, would not need to communicate with the ground—only with their neighbours, using laser links. They would therefore be able to devote much more of their mass to delivering processing power. And without the need for such accurate pointing, they could have solar panels that are slightly flexible, reducing their mass and further boosting specific power. Mr Musk has said that he thinks a specific power of 100W/kg is feasible for an AI satellite—and some believe that by using more efficient solar cells even 150W/kg may be possible in future. Philip Johnston, Starcloud’s boss, says his firm is aiming for a specific power of 70W/kg for its forthcoming satellites, based on what the company considers to be quite conservative assumptions. Moving on to satellite cost, Mr Johnston says Starcloud expects its design to cost “less than $5 per watt” when GPU costs are excluded. Mr McCalip estimates that Starlink’s current satellites cost around $22/W, down from $32/W for its original version. Again, an AI satellite should cost less to build, GPUs aside, because it does not require costly communications components. Move the sliders on Mr McCalip’s calculator to a specific power of 70W/kg and a satellite cost of $5/W, and the numbers look rather different: now the 1GW orbital data centre costs $16.7bn, only 5% more than the terrestrial one."

571

82 Comments

I’ve been working on the theory of why small, shippable factories are the perfect inversion to the centralized prefab factory model. It all collapsed into one line of math (credit to Gilles Retsin): Factory Efficiency = Usage × (Throughput × Product Value ) / ( CapEx + Deployment OpEx) The idea is that the closer you can drive Usage to 100 % while keeping the denominator tiny, the faster the flywheel spins.  A shippable microfactory tackles this by primarily focusing on a high Usage factor. Because the microfactory can be packed up afterwards and reused, its CapEx gets amortized over multiple projects, effectively raising its overall Usage across the year. If one project alone doesn’t fully occupy the factory, it can simply roll over to another job to maintain high Usage. Its CapEx is often an order of magnitude lower than a centralized factory (< $1M vs. $10-50M min), so the breakeven throughput is more achievable on a small pipeline of work. The real breakthrough here is portability. There is near-zero stranded capital and almost no idle time. To flesh this theory out, I wrote 3000 words on the topic for Brad Hargreaves and Thesis Driven. The article breaks down: - why “fixed factories, shipped goods” is being inverted to shipped factories, fixed goods - what a sub‑$1 M robotic cell does to CapEx per home versus a £45 M off‑site plant - early production/cost data on a shippable microfactory from Mollie Claypool, Gilles Retsin, Sam Baker and the Automated Architecture (AUAR) team - the hybrid reality of robots tackling the heavy structural components and human crews assembling faster and cheaper - technical constraints to pull this model off and future considerations on the optimal business model (own vs. rent) Major shoutout to Gilles Retsin for helping me pull this together. https://lnkd.in/eRTnbZPa

54

2 Comments

Thank Dani Cherkassky Great piece on the architecture shift voice AI needs. The core thesis: cloud-centric voice systems fail in real-world conditions because they are too slow, too expensive to keep always-on, and deaf to spatial context. The solution is a hybrid architecture—fast, always-on edge processing (Spatial Hearing AI and Cognition AI) handling 80% of interactions locally, with cloud LLMs reserved for complex reasoning. But the Kahneman thesis runs deeper than the article suggests. System 1 and System 2 are not just useful metaphors—they reflect an actual biological sequence. Fast, intuitive processing (System 1) evolved first; slow, deliberate reasoning (System 2) was layered on top. The brain's architecture reflects this: spatial-acoustic processing happens in milliseconds at the brainstem and primary auditory cortex, long before language centers engage. Kardome is not just borrowing Kahneman as marketing. Their architecture is neurologically correct—it replicates the hierarchical, parallel, spatially-aware processing that makes human auditory cognition so robust. Here is the deeper point: hearing preceded language in human evolution, and spatial hearing preceded both. The mammalian auditory system evolved to detect predators and locate prey—survival functions that required sub-second latency and continuous environmental awareness. Language processing was layered on top of this infrastructure. Current voice AI inverts this evolutionary logic: it starts with language models and treats spatial audio as an afterthought. Kardome restores the natural hierarchy. 🙌 Danny Shapiro, Rory Sutherland

35

2 Comments

The end of human drivers? Tried Tesla Robotaxi in the Bay Area for the first time, 8.8 miles from Meta SF to Daly City station for $18. Unlike Waymo, Tesla’s Robotaxi can take the freeway. The ride was smooth, and the decision-making felt far closer to human than “bot-like.” What surprised me most: it’s just a regular Model Y. With the built-in rear screen, any Tesla could potentially be activated as a Robotaxi. That’s scalability at an entirely different level. Waymo might have an early lead, but Tesla’s vision-only approach gives it massive coverage and faster rollout potential. If this scales, Uber, Lyft, and even car ownership itself could be disrupted faster than expected. This is exactly why I quit my job to go all-in on ElasticDash We’re entering a new industrial revolution, the real question is: will you shape it, or be forced to adapt?

15

2 Comments

While for Robot taxis, it is almost always ok to rely on the cloud, for real level 4 systems, you are going to be able to deal with the all O.D.D locally in the car, that means, having a local map always ready to take over, a lot of regulation of traffic understanding on board that Robot taxis don't actually need, as they use data telling them the most commonly driven path for places they know very very well. Level 4 is not Robotaxi, by far, and if you understand that a Level 4 car needs to deal with unknown roads, then you understand that you are going to pile multiple systems, dedicated to solve multiple different problems that robottaxi does not meet. The multiple systems on Chip of Athos Silicon was design to solve Level 4, and many other problems that requires multiple systems to work together without steeling the performance of each other, like most SDV systems are doing with VMs. The #mSoC is the solution if you have analyzed properly the problem you got to solve for really safe ADAS. https://lnkd.in/gSnFCpbM

12

Quantinuum’s planned IPO is the moment quantum finally hits the public markets in grown‑up form. This isn’t another early‑stage SPAC story; it’s a full‑stack platform spun out of Honeywell, with trapped‑ion hardware that regularly tops industry benchmarks and a real software and applications business on top. By combining industrial‑grade engineering with production‑grade tooling in chemistry, security, and optimization, Quantinuum gives investors exposure not just to “qubits,” but to an integrated quantum stack that already serves enterprise workloads. For technical and business leaders, as well as investors, the key question now is how public markets will price a quantum company that actually looks like an infrastructure and software provider rather than a speculative science project. Whatever the final valuation, this IPO will likely become the benchmark for what “mature” looks like in quantum—deep IP, differentiated hardware, and diversified revenue experiments before ringing the bell. Headline: Honeywell’s Quantinuum Quantum-Computing Unit Plans IPO – https://lnkd.in/euZjkGtG

194

5 Comments

On-device AI becomes powerful only when it works reliably on real hardware. In this Melange Walkthrough, Yeonseok Kim and Sun-Q Kim break down how the full on-device pipeline works, from model selection and quantization to real-device benchmarking and SDK deployment. Select. Benchmark. Deploy. This is how production-ready on-device AI should operate. Watch the full walkthrough below. #OnDeviceAI #EdgeAI #MobileAI #AIInfrastructure #ZETIC #Melange

26