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

The AI Market Looks Different From a Back Patio A lot of people are questioning AI right now. Investors and analysts have spent months asking whether the enormous investment in AI infrastructure will pay off, pointing to signals like Copilot adoption rates and enterprise deployment metrics as evidence that real demand might be softer than the capex implies. Every earnings season brings a fresh round of debate about whether businesses will actually pay for this technology at scale. Last night I had an experience that speaks directly to that question. I was sitting around a patio in the Midwest with a mix of old friends and new acquaintances. The conversation eventually worked its way to the question every group of adults lands on, what does everybody do for a living. I told them I test and benchmark AI systems, and that answer took over much of the night. Every single person on that patio uses AI prolifically in their day-to-day work. One works in finance, another in M&A, another in manufacturing, one in IT, and one in sales. Not a software developer in the group, and this was nowhere near Silicon Valley or any other tech hub. Everyone there agreed on two things. The first is that AI has become genuinely important to how they work. It makes them significantly faster and more effective, and they know it. These are not people evangelizing a technology because their company sells it. They found tools that work, built them into their routines, and kept using them because the results were obvious. The second is that every one of them acknowledged a stigma around admitting it. In some of their workplaces that stigma still exists. In others it existed until recently. The pattern was consistent either way. People are using AI to do their jobs better and staying relatively quiet about it. That second point goes straight back to the adoption debate. If professionals in finance, M&A, manufacturing, IT, and sales are using AI every single day but hesitant to advertise it at work, then surveys, seat counts, and official deployment numbers are missing a meaningful slice of actual usage. The adoption is happening whether or not it shows up cleanly in a dashboard. Nobody on that patio was debating capex or token economics. They were just getting more work done. What surprised me is that a couple of them were genuinely curious about where the compute behind all of this happens, whether the AI they rely on runs locally on their device or in the cloud. When people with no connection to the technology industry start asking where the compute lives, the conversation has moved well past the early adopter phase. I spend my professional life benchmarking AI systems and infrastructure at @Signal_65, so I see a lot of data on this market. Sometimes the clearest signal does not come from a lab or an earnings call. It comes from a back patio in Kentucky, with people far removed from the industry all telling you the same thing without coordinating their answers. This is a giant market, it is still early, and it is not going back. Picture of dog on above mentioned patio for likes.

Match the Mac, then do more. Beyond the benchmarks, @Windows 11 covers ground the Mac lineup does not: PC gaming, touch and pen 2-in-1s, displays from 11 to 17 inches and OLED, full-size connectivity, and real choice across price points. Same categories, more capability. Read the report: signal65.com/research/measu…

This is the argument I found most compelling coming out of the Qualcomm Investor Day last month. The Dragonfly story is not really about any one single accelerator or processor, but more about the memory architecture underneath the roadmap. Qualcomm cites estimates that transformer models grew 240x every two years over the past decade while AI hardware memory grew just 2x. High Bandwidth Compute (HBC) moves compute into the memory package itself, so the internal bandwidth of the DRAM stack becomes the effective bandwidth available to the workload rather than whatever it is on the trip across an external interface. The roadmap claims are aggressive - that HBC Gen 1 on Dragonfly AI250 is designed to deliver an 18x increase in effective memory bandwidth over AI200, and Gen 2 on AI300 targets 54x. Those are projections, not measured results, and delivered performance on real inference workloads is what will settle this debate. That validation work is squarely where @Signal_65 will spends its time. If the economics of inference are set by data movement, the winners could be the architectures that shorten that journey. Qualcomm is making that bet now, and the industry will be paying attention.

The assumption that Macs are simply faster no longer holds the way it used to. In head-to-head lab testing, a mainstream @Windows 11 PC matched the MacBook Air M5 on single-threaded and creative work, and pulled ahead where it counts, in our testing: ✅ +92% multi-thread CPU (Cinebench 2026) ✅ +78% on-device AI (Procyon AI, NPU) ✅ +50% Microsoft Word Full report: signal65.com/research/measu…

When a small team shares one AI appliance, the number that matters is not just single-user speed but how the system holds up when several people query it at once. We ran the @nvidia DGX Spark as a RAG chatbot on a 30B model across 1, 2, 4, and 8 concurrent users to find out. Per-user throughput dips under load, but total system throughput climbs about 4x as the vLLM continuous batching interleaves the request streams rather than serializing them. Full case study: signal65.com/research/ai/ra… What we found in our testing: ➡️ Single-user throughput ran about 56 tokens per second, tapering to about 28.5 tps at 8 users, still above the comfortable reading rate ➡️ Aggregate throughput rose from about 56 tokens per second at 1 user to about 228 tps at 8 users, roughly 4x more total work under load ➡️ 4 concurrent users saw a 7.4s time to first token at about 39 tokens per second each, a responsive experience for interactive chat ➡️ Latency and throughput taper gradually rather than collapsing, so a team can grow into the appliance and sense the limits before they turn disruptive

Futurum AI 15 rebalance coming soon. 👀👏🏻

Cobalt 200…

The desktop has powered corporate productivity for decades, yet the AI PC conversation has stayed focused on laptops. Our new study asks whether AI changes the math on the desktop, and whether the silicon underneath the workload actually matters. Full report: signal65.com/research/ai/me… Running a common set of office tasks on an HP EliteDesk 8 SFF with the @AMDRyzen AI 7 PRO 450G, AI assistance cut total time to 19.1 minutes from 56.7 minutes done manually, a 66% reduction in our testing. Across the same testing: ➡️ Individual workflow steps ran as much as 17x faster with AI, with meeting summarization showing the largest gain ➡️ At typical task frequencies, AI users save roughly 1.1 workdays per week, about 11 work weeks per year ➡️ The current-generation chip ran the same AI workloads 1.8x faster than a prior-generation Ryzen desktop ➡️ On some workflows the older desktop ran AI slower than the manual approach, a sign that modern silicon is required to make these workflows worth running at all

Small teams want production-grade AI without a cloud subscription, a server room, or a dedicated infrastructure hire. We ran the @nvidia DGX Spark through a full day/night cycle, serving RAG inference for the team by day and fine-tuning a model with LoRA overnight, to see whether one desktop can cover both halves of a real AI deployment. Full case study: signal65.com/research/ai/ra… What the case study found: ➡️ In our testing, 4 concurrent users saw a 7.4s time to first token at roughly 39 tokens per second each, with total system throughput rising about 4x going from 1 to 8 users ➡️ In our testing, overnight LoRA fine-tuning held about 425 tokens per second at steady state, finishing a 500-step run in roughly 45 minutes ➡️ 128 GB of unified memory serves a 30B model in NVFP4 by day and fine-tunes a model at night, with no changes to the model itself ➡️ At $4,700 across 8 users, the hardware works out to under $600 per seat, a one-time cost with no cloud meter and no data leaving the building

$QCOM investor day this week. I expect a few positive datacenter surprises and a much larger TAM for it to address. 👀💪🏻

The AI industry has a benchmark problem. Training contamination, LLM-as-judge guesswork, and gameable test sets have produced leaderboards that tell enterprises almost nothing about real-world retrieval. Full report: signal65.com/research/ai/be… So we partnered with @KamiwazaAI and built one that fights back. RIKER generates a fresh, contamination-resistant document corpus with deterministic answer keys and zero human annotation, then grades retrieval and hallucination against ground truth it fully controls. We ran 91 models through it. The results should make every vendor selling on context-window size nervous. ➡️ 27 models cleared 95% accuracy at 32K context. At 200K, only 3 survived. ➡️ Multi-document aggregation accuracy collapsed more than 2x faster than single-document retrieval. ➡️ Bigger context windows did not mean better retrieval. The token-count arms race is marketing, not capability. ➡️ Thinking models boosted retrieval by up to 64% and swept every top spot in our testing. ➡️ Hallucination held steadier than retrieval as context grew, a failure mode most evals never even isolate. Advertised context length is not a proxy for usable retrieval. If your AI strategy assumes otherwise, read this first.

Lots of hands on time with early @NVIDIARTXSpark systems today. First up, the @Dell XPS 16 Creator Edition. Very familiar, high quality design with a look and feel of a premium device for this audience. (Note no partner had systems available to turn on outside NV demo room.)

During the @NVIDIA Q&A with Jensen I had the chance to ask the final question: "With the PC market being low margin and cut throat, why enter it now?" His answer was compelling. He dismisses margin concerns and focuses on the opportunity to add value and reinvent what we love.

I walked out of the @Qualcomm keynote at Computex this week convinced the milliwatts-to-megawatts strategy is the most coherent vision in compute. The one piece missing was the concrete next step. Two days later at Microsoft Build, Project Solara answered it. A chip-to-cloud platform for agent-first devices, with @cristianoamon sharing the stage with @satyanadella, and a wearable badge that lands at the low-power end where Qualcomm has spent decades building an advantage. The full vision still runs through the data center, and that gets earned through execution, with more news landing June 24. What Solara confirms is that the personal-device franchise is a real strength to lean on. Full take linked below.

NVIDIA Is Coming to the PC. Watch Who Suddenly Becomes an Expert. Over the next few days you are going to see a lot of people on social media "discover" the PC. They will have strong opinions about the PC ecosystem, about AI on the PC, about the engineering tradeoffs behind a thin and light notebook, and most of them will be forming those opinions for the very first time. The reason is simple. @nvidia now has a voice in this conversation, and that voice pulls a crowd. Now that doesn't change the fact that this is a real moment. NVIDIA is the 5 trillion pound gorilla in the room, and when a company that size leans into a market, things move. I am genuinely excited about what the N1X and the family of chips around it could bring to the PC, and about how it might push @Windows, and @Microsoft more broadly, in a new direction. But excitement is not the same as suspending the rules. I do not believe, as some of the more out-there posts are already claiming, that NVIDIA is about to make every laptop on the shelf irrelevant overnight. Silicon is still bound by physics. Power, heat, and thermals translate directly into performance and battery life, and there is no way around that. If this turns out to be a GB10-based part as rumored, then there is no magic here that we have not already measured. This includes the @MediaTek CPU based on @Arm and the NVIDIA GPU itself. We at @Signal_65 put the platform through its paces in our DGX Spark work, and the numbers are the numbers. signal65.com/research/ai/th… Silicon is also bound by the software around it. Windows 11, Windows on Arm, application compatibility, and gaming support all sit between a great chip and a great experience. That layer does not bend just because a new logo walks into the room. So does NVIDIA have the engineering ability, and the political weight in this industry, to move things that have been stuck in the mud for years? Absolutely. That is the part I find most interesting. Here is what that could actually look like. Windows on Arm could finally be treated as a first class citizen. Gaming on Windows on Arm could get a real leg up, because NVIDIA can compel developers to get involved in a way that few others can. And Windows itself could be pushed further into the AI era, beyond the current Copilot+ features, and grow into a leading AI development platform. Does any of this make Intel, AMD, or Qualcomm irrelevant? Far from it. Intel is getting its legs back under it with Panther Lake. AMD keeps iterating on high performance designs like Strix Halo. And Qualcomm, along with Snapdragon, arguably benefits if NVIDIA is now pulling for Windows on Arm too. A bigger tent helps everyone building inside it. But remember, be a little careful with that "new" crowd. A lot of the same people will also hand you confident, "informed" views on the data center. Ask yourself whether the take in front of you comes from someone who has actually done the work in this space, or from someone who only got interested because NVIDIA showed up. Those are not the same thing, and the difference matters more right now than it usually does. This week is going to be very, very interesting. My ask is simple. Balance the excitement with the right questions and the right thinking. A new entrant in this space is going to be good for the consumer and I think it is finally time for a reset of what we mean by the term "AI PC." See you this week.

At Computex, Qualcomm is introduced @Snapdragon C, a new entry-tier platform built for Windows laptops targeting around $300 (and up), with Acer, HP, and Lenovo signed on and the @Acer Aspire Go 15 leading as the first device. The positioning is the obvious headline, but an interesting story sits in the architecture. Snapdragon C steps away from the custom Oryon cores that define Snapdragon X and X2 and uses a mobile-derived Kryo design instead. That is the same fundamental move Apple made by dropping an iPhone class A18 Pro into the MacBook Neo. Both companies are taking proven, high-volume phone silicon downmarket to reach a price the flagship laptop architecture was never built to hit. Snapdragon C is the first Snapdragon PC platform that does not support Copilot+. It still carries an integrated NPU for everyday on-device AI, but I assume it sits below the 40 TOPS threshold that the X series (45 TOPS) and X2 Elite (80 TOPS) clear and its unclear if Microsoft will keep the 16GB memory requirement too. How this translates into "on-device AI" will be an interesting question this summer. The lineup reads as a clean three-tier stack. Snapdragon C anchors the value segment, the X series covers mainstream premium now as the previous generation, and X2 Elite leads at the top. It is also worth being precise about who C actually competes with. At $300 and up, the more direct fight is less the MacBook Neo and more the Chromebook and budget x86 field, namely Intel N-series, MediaTek Kompanio, and AMD in the entry tier. The question is price. Chipmakers do not set laptop prices, OEMs do, and the current memory and storage shortage makes a genuine $300 sticker hard to deliver right now. The strategy is sound and the segment is real, but the number that matters is where devices like the Acer Aspire Go 15 actually land on the shelf. That is what will tell us whether Snapdragon C reaches the audience it was designed for.

Hmmm, I wonder... 😂

I've really enjoyed our time working with @Lenovo exploring their AI solutions stack and how they actively work with enterprises on deployment, not only getting to a concept state. This is a good discussion that summarizes one of our recent @Signal_65 papers. signal65.com/research/ai/le…

🤝 @TheFuturumGroup has entered into a definitive agreement to acquire @ETRnews. This announcement marks an important step in Futurum’s continued focus on decision-grade intelligence for technology leaders, investors, and enterprise decision-makers. By bringing together Futurum’s analyst expertise, advisory, intelligence, and strategic content capabilities with ETR’s institutional-grade technology spending data, this combination is designed to strengthen the signal organ Hear directly from @danielnewmanUV and @bmlascolea on what this means for Futurum, ETR, and the future of technology intelligence. #TheFuturumGroup #ETR #TechResearch

The 2026 @Windows refresh story isn't a marketing claim, it's a measurable performance gap, and it's a large one. We tested the AMD Ryzen AI 9 HX 465, Intel Core Ultra X7 358H, and Qualcomm Snapdragon X2 against a representative five-year-old laptop running Tiger Lake. In our testing, every 2026 platform delivered between 3.7x and 7.2x the multi-thread CPU performance, with comparable gains in graphics and content creation. For the 340M PCs sold in 2021 now entering their refresh window, this is the upgrade story. Full report: signal65.com/research/insig…