Signal65

We’re excited to launch @Signal_65, a new company that will serve the technology industry with testing, performance validation, and data-based consulting. @ryanshrout will serve as our President and GM, with @PatrickMoorhead & @danielnewmanUV on our board. prn.to/47Omv1C

That claim @MichaelDell made from the @Dell Tech World keynote stage, breaking even on agentic AI versus cloud APIs in as little as three months and reducing spend by up to 87% over two years, comes from a study my team at @Signal_65 published this week. Full analysis here: signal65.com/research/ai/th… We modeled agentic workloads across general knowledge, sales, and software development on a five-day work week, with publicly available API pricing on one side and Dell workstation and server pricing on the other. Savings scaled with concurrency and model size, with the strongest economics coming from workhorse models in the 30B to 284B parameter range, which is where the bulk of useful agentic reasoning actually runs today. That is the financial backdrop for one of the most important Dell announcement of the day, Dell Deskside Agentic AI. This is the first time I have seen a major OEM really drive agentic AI on the desk, in the rack, and across the data center under one consistent runtime and security model. @NVIDIA OpenShell now spans the whole Dell AI Factory with NVIDIA, from a Dell Pro Max with GB10 up through PowerEdge XE servers. That continuity is what gives enterprises a path from prototype to production without rebuilding the stack. The strategic backdrop is the data foundation. Updates to the Dell AI Data Platform around orchestration, search, and a Starburst-powered SQL engine accelerated on NVIDIA Blackwell change how enterprise data gets fed into AI pipelines. PowerRack now brings block, file, and object storage onto one rack architecture with PowerFlex joining Exascale. Dell is making sure that whatever infrastructure plan you have, they can address. Google Gemini 3 Flash on Distributed Cloud running on PowerEdge XE9780. OpenAI Codex connecting to the Dell AI Data Platform. Palantir Foundry coming on-prem. Reflection and SpaceXAI models landing on Dell infrastructure. Dell Enterprise Hub on Hugging Face expanding to MiniMax-M2.7, DeepSeek V4, GLM 5.1, and Kimi K2.6. Whatever model an enterprise wants to run, Dell wants the on-prem stack to be the answer. The customer evidence is already in production. Mistral AI training on liquid-cooled PowerRack with NVIDIA GB200 NVL72. Eli Lilly feeding more than 1,000 GPUs at nearly two terabytes per second on LillyPod for drug discovery. Samsung embedding Dell infrastructure across global semiconductor fabs. Cloud-only was going to hit an economics wall as agentic workloads scaled token consumption. The more interesting question now is how fast enterprises move, and which workloads they bring on-prem first.

At CES, @AMD pulled the cover off the Ryzen AI Halo, a first-party mini PC built on the Ryzen AI Max+ 395 processor and aimed directly at the NVIDIA DGX Spark. Now we know it is priced at $3,999 with 128GB of unified memory and a 2TB SSD, with pre-orders expected next month (June). AMD also confirmed the next generation, the @AMDRyzen AI Max PRO 400 Series, code-named Gorgon Halo, as the follow-on platform. The market AMD is chasing, local AI development and small-scale on-device inference, is only just forming. The prize is owning the developer on-ramp, the desktop box where builders prototype, fine-tune, and run models without renting cloud capacity. When a category is this young, a second credible vendor matters. And AMD is showing up as a real competitor, not a token entrant. Shipping a first-party platform rather than leaving it to partners is a signal of intent. So is supporting both Windows and Linux with full ROCm. In a space where every developer workflow looks a little different, meeting people where they already work is a smarter early bet than it looks. What about performance? On paper, the Ryzen AI Halo and DGX Spark are closely matched, and the early AMD numbers against Spark look strong. But those are vendor claims on a short list of models, and they need independent validation. Winning a benchmark is the easy part. The real test is performance across a wide and constantly shifting set of models, plus staying current as new models, workflows, and inference stacks land almost every week. That ongoing support is where the @NVIDIA CUDA ecosystem has a long head start, and it is the bar AMD has to clear. Why does any of this matter in 2026? Agents. Autonomous agents run continuously and consume far more tokens than chat-style interactions. Recent Signal65 research (signal65.com/research/ai/th…) on the economics of agentic AI found agentic workloads burn roughly 4x to 15x more tokens than standard chat, and the trend points well beyond that. When consumption scales like that, per-token cloud pricing becomes a real line item, and owning local inference capacity stops being a hobbyist choice and starts being an economic one. That brings us to the Ryzen AI Max 400. The CPU and GPU look like a measured step over the current part, so memory is the key difference. Moving up to 192GB of unified memory is the spec that counts, because memory capacity decides which models you can actually fit and run locally. That is a significant jump over the 128GB ceiling on both the current Halo and the DGX Spark. AMD is aiming for the third quarter of 2026 for the Ryzen AI Max PRO 400 platform, so the bigger unknown is price. The first-party Halo at $3,999 undercuts the $4,699 Spark, though not by the wide margin some expected, and Ryzen AI Max 400 pricing has not been disclosed. In the current memory market, that is a genuine variable, and hitting that third-quarter window is not a given either. Local AI is shifting from a curiosity to a real platform decision, and AMD just made it a two-horse race. Are you planning to run inference locally in 2026, or is the cloud still the default for your team? I would like to hear how you are thinking about it.

Most enterprise AI deployments are still cloud-first. That makes sense for bursty workloads, but persistent agents could change the calculation. When inference runs continuously, infrastructure pays itself back faster than most procurement timelines assume. Across the @Dell AI Factory with @NVIDIA portfolio, we modeled breakeven against equivalent cloud API spend at every tier: ➡️ Dell PowerEdge XE7745 with NVIDIA H200 NVL GPUs, 2 months ➡️ Dell T2 Workstations with NVIDIA RTX PRO 6000 BW, 2 to 7 months ➡️ Dell Pro Max with NVIDIA GB300, 3 to 11 months ➡️ Dell Pro Max with NVIDIA GB10, 6 to 17 months Every month past breakeven is cloud spend you aren't paying. Full report: signal65.com/research/ai/th…

$NVDA CEO Jensen Huang to $DELL CEO Michael Dell “We have now entered the era of useful AI.” Reiterating my endless droning on about just how early it is for AI. We have literally just hit the point where it is beginning to do useful things. So. Damn. Early. 👏🏻🚀

@GeminiApp @OpenAI @grok @SpaceX @nvidia The play for deskside AI compute could be a massive win for Dell as the need for more on-prem processing expands. @MichaelDell quoted our @Signal_65 “break even in 3 months” claim for GB300!

The economics of persistent AI agents look different from chatbot workloads. Always-on agents consume orders of magnitude more tokens, and per-token cloud pricing isn't built for that pattern. Our report: signal65.com/research/ai/th… The most striking case in our analysis came from software development workloads. One @Dell Pro Max with @nvidia GB300 workstation delivered 87% lower cost than the equivalent cloud API spend and saved $926K over two years. That is one workstation, under a desk, doing the work of a $1.06M cloud bill. The full report covers two more workload profiles and scales up through the Dell workstation portfolio to PowerEdge.

Persistent AI agents don't consume tokens like chatbots do. They run continuously, generate at least 4x to 15x more tokens than single-turn interactions we've gotten used to, and autonomous agents may push that closer to 1000x. It fundamentally changes the math on token production, and on-prem compute vs cloud APIs. We modeled three persistent agent workloads on @Dell AI Factory with @nvidia infrastructure vs. cloud over a two year period. What we found: ➡️ On-prem reduced costs by 28% to 90%+ across all workloads ➡️ Dell Pro Max with NVIDIA GB300 Ultra delivered 87% lower cost for software development, $926K in two-year savings from a single workstation ➡️ Most platforms broke even in under a year, some in as few as 2 months ➡️ The portfolio scales from desktops handling 8 agents to PowerEdge systems supporting 18,000+ Full report: signal65.com/research/ai/th…

An interesting and critical topic for enterprise and a great conversation. Thanks @HPE_Compute

Arm reported Q4 FYE26 results are strong, but the forward signal that is most interesting is in the data center business with AGI. Revenue of $1.49B, up 20% YoY, the highest quarter @Arm has ever had. EPS of $0.60 beats consensus. Q1 FYE27 revenue guide of $1.26B goes above the street estimates. The bigger story is the Arm AGI CPU, which now has more than $2 billion in customer demand across FY27 and FY28. That figure has more than doubled since the launch announcement just 5 weeks ago. The royalty side is moving in lockstep with data center royalties more than doubled YoY. Arm now holds roughly 50 percent of CPU compute share among top hyperscalers (based on Arm statements). Google announced Axion hosts in the new TPU8t and TPU8i, NVIDIA unveiled Vera at GTC, and AWS reported its custom silicon business, including Graviton, at a $20 billion+ annual run rate. Arm has crossed from IP licensor to vertical compute platform. The competition from x86 to Arm in AI infrastructure is no longer a forecast, it is the operating reality of the hyperscaler buildout. Curious how investors are weighing the AGI CPU ramp against the platform royalty engine. Both vectors are working at the same time and that combination is rare in this market.

In our latest Lab Insights report, commissioned by Microsoft, Signal65 found that @Windows 11 PCs starting $150 below the MacBook Neo delivered more performance AND up to 56% longer battery life in Procyon Office Productivity testing. Performance and endurance are usually framed as a tradeoff. Our testing showed Windows buyers don't have to choose, and eligible US college students can layer on a $500+ Microsoft bundle that adds even more value. Read the full report for benchmark methodology and system configurations. signal65.com/research/windo…

I've been listening to the analysts and leaders at silicon companies talk about agentic AI driving more CPUs in the CPU:GPU ratio, and it's something we have seen coming for quite a while now, even in our own @Signal_65 testing. But something to keep in mind 1/4

Micron just started shipping the 6600 ION at 245TB, the highest-capacity commercially available SSD on the market. Wow. Compared to HDD deployments at the same capacity, it requires 82% fewer racks and runs at roughly half the power. On AI workloads, Micron's own lab testing shows up to 84x better energy efficiency, 8.6x faster preprocessing, and 29x lower latency. Object storage gains are even larger. (Obviously we are looking to get these claims run from @Signal_65 as well.) Power has become the constraint on AI infrastructure scale, and drives like this are exactly why the data center storage mix keeps shifting away from HDDs. AI data and large-scale object storage are the obvious near-term targets, and Dell is already integrating into PowerEdge. Can I get one for my gaming rig too??!?

Bringing in @AlexKatouzian to lead the client group and to jump start physical AI at @Intel is a big move, and will be one with impacts that range to areas we haven't seen yet. Alex did great work managing business and technology at Qualcomm and its partners, looking forward to seeing what we can do with Intel

The online narrative would leave you to believe there is only one option for a high-quality, performant laptop that anchors in value. In our latest Lab Insights report @Signal_65 measured up to 92% faster CPU performance from @Windows 11 PCs versus the MacBook Neo at matching (or better!) price points. Across four mainstream laptops, Windows systems also led in Adobe Photoshop, Microsoft Word, and Office Productivity, all while remaining eligible for a $500+ Microsoft college student bundle that adds Game Pass Ultimate, Microsoft 365 Premium, and an Xbox controller. Read the full report for the benchmark data and configurations. signal65.com/research/windo…

The memory wall has been the limiting factor for HPC for years. Core counts kept climbing, but DDR5 channels could not keep up. The new @Azure HBv5 instances change the geometry. By placing HBM3 stacks directly on the custom @AMDServer EPYC package, sustained memory bandwidth climbs from roughly 780 GB/s on HBv4 to 6.9 TB/s on HBv5. A 9x increase. For memory-bound workloads, that is the difference between cores running at full saturation and cores stalling on data fetches. Full report: signal65.com/research/a-tec…

Edge AI gets more interesting once the environment stops cooperating. @Signal_65’s @ryanshrout and Russ Fellows talk with @HPE’s Vincent Sheu about the DL145 Gen11’s inference performance, and what actually matters when AI leaves the data center.

For HPC workloads, time-to-solution is the metric that matters. Faster iterations mean more design cycles, earlier forecasts, and shorter research timelines. In our testing, the new @Azure HBv5 instances completed an OpenFOAM CFD run with a 100M-cell mesh in 261 seconds. The prior HBv4 generation took 1,278 seconds. That is a 4.9x speedup on a single workload. The driver is integrated HBM3 keeping all 368 cores on the @AMDServer CPU fed with data, not waiting on DDR5 channels. Full report: signal65.com/research/a-tec…

AI platform architecture What does the future of AI infrastructure look like at the platform level? Ryan Shrout from @Signal_65 joins AMDers Curt Waltman and Madhu Rangarajan to discuss how CPUs, GPUs, networking, and open ecosystems work together to scale modern AI systems. Read through the Signal65 research paper for more insights: signal65.com/research/ai/im…