@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20261005

Episode Date: October 5, 2026

- Tencent on Oracle GPU cloud - Autowarcom and Edge-HPC - Google Suncatcher gets feet wet - HBM further eats into DRAM wafer budget [audio mp3="https://orionx.net/wp-content/uploads/2026/10/HPCNB_20261005.mp3"][/audio] The post HPC News Bytes – 20261005 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HPC Newsbytes. I'm Doug Black, and with me is Shaheen Khan. Reuters reports that Chinese company Tencent has agreed to lease access to roughly 100,000 advanced AI accelerators from Oracle in a five-year deal worth about $7 billion. processors remain outside of China in Oracle data centers across Southeast Asia, so a Chinese company restricted from buying some of the world's most advanced processors can potentially rent the computation offshore. That's the architecture we've discussed before.

On-prem can mean domestic while cloud can mean offshore. Export controls still constrain China's ability to build unlimited amounts of advanced compute domestically. but if offshore cloud access remains available, the control merely restricts ownership and location rather than access to the computation. Supposing that path would be considerably difficult to enforce and would involve allies, cloud providers, and cross-border data policy.

Yeah, it also marks a distinction between control of the infrastructure and control of the model and the data. And this is a good segue into, all the models that are out there, I think it's time to take stock of what that's seen is internationally. At the organizational level, the United States has Anthropic, which is noted for coding and agents, Google Deep Mind for multimodal and science. I mean, they all do everything, but this is kind of what comes up as their visible strength. Open AI for reasoning and agents, space X slash XAI, also for coding and agents, meta, pursuing a more open ecosystem approach, Microsoft AI, which increasingly has its

own models across reasoning, voice, vision, and agents. They aren't all sitting at the same point on the frontier, let's say, but they're all independently building substantial model technology. If you're not tracking China, you might be surprised with the breadth of the field there. Quinn from Alibaba with broad and open weights, Deepseek, which we all have heard, of focused on efficiency and open weights, Minimax, noted for agentic and low-cost operations, Z.AI, they have a system called GLM for open reasoning, Moonshot and Kimi for long contacts and agents, seed from byte dance for multimoral agents, 10 cents, Hun Yuan for full-stack deployment, and Baidu's Ernie, the domestic platform ecosystem. In Europe,

France's mistral remains the most obvious general purpose contender, with a focus on openness and efficiency and subject to European data regulations too. Then there is another category that I wouldn't call frontier LLMs, but which could be strategically important. France's AME labs, pursuing a novel strategy called world models that could be the next big thing. Canada's cohere, driving enterprise and sovereign AI. Korea's neighbor focused on sovereignty and Korean language models. Korea's LG AI research has something called X-A-1, focused on industrial and scientific AI.

The UAE has Falcon, working on compact reasoning. Japan's Sankana, noted for model evolution and orchestration. India's Sarvam, with a focus on Indian languages, and Israel's AI-21, driving hybrid architectures. You could say that below these largest, models, there is another interesting race in small language models that might find their way into bigger AI systems. Microsoft is pursuing fewer bits with the 1.58-bit bit bit net, and also its five family of small models and highly optimized computer use agents with FARA, FARA. The US-based

Prism ML, we've covered them here, a Caltech spin-off, is going further with a one-bit model. Google has gamma also for smaller local deployments. UAE's TII is for compact reasoning models, and Canada's cohere also has agent decoding models with only a few billion active parameters. It's also noteworthy that for technology as complex as AI, there are so many players globally. So important differences are going to be a level or two deeper,

which also means the value of having the single highest scoring model is gradually declining. The big walkaway for me is this. Once there are 10 or 15 models that are sufficiently capable, competitive advantage moves towards proprietary data, domain expertise, application integration, deployment economics, and the ability to choose the right model for the right task. Pentagon plans to create something called the Autonomous Warfare Command or Auto-Warcom

focused on autonomous and robotic systems across, the military. It is scheduled for Q4 2027, although some of those authorities require congressional approval. Auto-Worcom would be another combatant command, but organized around a function, which is autonomous warfare, rather than a geographic region or physical domain. We'd be interested in the computing architecture behind this mission. HPC and cloud systems trained models, run simulations, generate synthetic environments, views, intelligence, and update fleets. At the other end, you have HPC at the edge. Transportable supercomputers, drones, vehicles, ships, and robots need local perceptions, sensor fusion, navigation, and decision-making. And they cannot depend on continuous network access.

Army planners use the term DDIL, denied, disconnected, intermittent, and. and limited communications. In a conflict, networks can be jammed, attacked, or simply unavailable. So enough AI has to travel within the platform for it to continue operating locally. Very interesting. And another piece of evidence of how technology drives organizations,

workflows, and ultimately culture. As physical AI pushes HPC technologies out of the data center and into systems that move around the physical world, it creates possibilities that demand new strategies and actions. And the Pentagon increasingly seems to be treating it as a persistent operational need closer to cyber or space rather than another technology program. Every device, everything, really, wants to be an AI-enabled computer. A car is a computer on wheels, as they say, but every wheel is a computer too.

So drones can have enough computation for immediate perception and control. A mini-supercomputer mounted on a vehicle or at a command post can aggregate data and coordinate multiple systems. Theater-level edge-hPC clusters can do larger inference planning and model updates. The national supercomputers and cloud facilities could do the really large training simulation and intelligence jobs. When connectivity disappears, the lower layers can continue, but they need lots of storage to have the necessary data and to queue up some things. But anyway, continuous external connectivity becomes something that improves the system, rather than being a show-stopping gating factor. And space fits directly into that hierarchy, supplying communications, positioning, sensing,

and increasingly autonomous spacecraft. I know there will be a lot to work out and details that make or break things, so it's really complex, I know, but the high-level sketch is materializing. The broader and interesting point is how EdgeHBC is becoming a military requirement. Google has put AI accelerators in orbit. Its first project Suncatcher satellite launched October 1st on a SpaceX Rideshare mission. It went up as one of 130 payloads sharing the same rocket. The spacecraft is described as roughly refrigerator-sized and contains four of Google's TPU accelerators.

The attraction is solar power. In the low Earth orbit, Google calculates that solar panels can produce as much as eight times the energy of equivalent panels on Earth because they've received the suns rays for more of the day and are not blocked by clouds and atmosphere. But power immediately turns into the problem of heat. There is no air in orbit, so you can't blow air across the chip or use liquid cooling. Google's path is a TPU to thermal interface into copper and aluminum heat pipes and then into an external radiator, which emits the energy away as infrared electromagnetic radiation into space. Google's longer-term architecture has satellites carrying dozens of accelerators and communications through high-bendwidth laser links.

Its design studies include a cluster of 81 satellites flying within a one. kilometer radius. The economic model also makes a strong assumption, launch cost to low Earth orbit eventually falling to around $200 per kilogram or below by the mid-2030s. It's great progress and a logical next step from HPE's project years ago, which demonstrated that fairly conventional HPC hardware can operate in space. Their so-called space-borne computer was on the International Space Station, which supplied power, a controlled environment, and a station-wide thermal management system with large external radiators,

and the system was small enough to be absorbed into that setup. Google's system sounds like a tiny one, but Google has more hardware in orbit, a peer-reviewed system study, radiation testing, a thermal experiment, and follow-on optical interconnect experiments planned. It's small scale, but definitely a step up. But putting four processors in order, orbit for local use or specific mission is a radically different engineering problem than create an AI

factory as a cloud service for users on Earth. AI racks are approaching a megawatt per rack, and AI factories have tens to hundreds of such racks. So for the sake of argument, if a satellite consumes a megawatt, then about a megawatt eventually has to leave as heat. A gigawatt scale system eventually has to dispose something close to a gigawatt of heat. all of that through radiation. So converting heat to radiation will be the key technology. And there are a few startups working on that.

A highly distributed system with a relatively low-power system on board could simplify things and could standardize into higher volume modules and might be the way to go. But our current notions of an AI factory is not on the cards anytime soon. Google wants satellites moving independently through orbit to behave like a tightly coupled compute cluster using laser links. That's possible in principle, but the bandwidth, synchronization, pointing accuracy, and reliability requirements are substantial. Then there is the mass equation. Radiators, solar arrays,

optical terminals, shielding, and structural hardware all have to write the rocket along with the processors. Google's own economics look interesting only if launch prices fall dramatically from today's levels. So I would put it where I put a lot of advances these days. Great progress and a great distance to go. It has a prayer, especially for local workloads originating in space, satellite imagery, sensor processing, and local processing and data reduction before sending data back to Earth. But no AI factories as we know them, despite how enticing access to lots of solar energy is. A quick memory update, Samsung says, I've been with memory, could consume nearly 30% of global DRAM wafer capacity in 27, up from about 20%.

Today, HBM and ordinary DRAM ultimately compete for a substantial amount of the same manufacturing capacity. So manufacturers have to balance how much of their capacity to allocate to each. and there is a distinction between wafer capacity and bits produced. HBM uses large dyes and more complicated stacking and processing, so it consumes fabrication capacity disproportionately. Yes, the analyst's firm Trendforce has arrived at roughly the same conclusion, estimating that HBM could consume around 30% of wafer input while producing only about 13% of DRAM bits.

Trendforce says the pressure is already strong enough that accelerator designers are considering lower capacity 8 high HBM configurations for some future products. And of course, some inference chips use static RAM. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X. Shaheen Kahn and Doug Black host the show.

episode is posted on Orionx.net. If you like the show, please rate and review it. Thank you for listening.