ARM’s 2026 ‘Tech’ Day Is A Self-Inflicted Wound

How could they get so bad so fast?

ARM LogoARM’s most recent ‘Tech’ Day was an embarrassment for the company, destroying a reuptation for solid messaging built over a decade. While there might have been one, SemiAccurate can’t think of a worse briefing in the 20+ years we have been covering the industry, it was that bad.

Before we go into the insipid waste that ARM messaging has become, lets kick things off with a little history lesson. Up until the early 2010’s, ARM didn’t talk much about anything to anyone other than partners. They had an attitude that if they talked, they would step on the toes of their partners. Partners in turn wanted to talk about everything but the licensed ARM IP, that bit is understandable. So ARM’s products never got any technical press because there was really no way to get the information.

On October 11, 2012, this deadlock took it’s first steps to being broken with an email from the author to Simon Segars, ARM CEO at the time, explaining the issue and asking for changes. This lead to some calls with PR, and on May 17, 2013 the first ARM Tech Day happened. It was a small event for which the slides are no longer in my possession but some memorable pictures were taken for use a year later. (Note: The date in the linked story is wrong, the EXIF date on the pics is as above)

The first two years were deeply technical, that was the point of Architecture/Tech Day after all. There are a small number of press out there that care about technical details so you get them in a room and give them the things they need to do their job. It isn’t that complex a paradigm and it worked well. The first two years of slides have evaded my search skills but year three, 2015, had deeply technical slides like this one.

ARM 2015 CPU technical slide

This is more tech heavy than 2026 slides

Umm… like this one.

ARM 2015 CPU technical slide A72 uarch

ARM A72 architecture slide

Since then the author has attended every Tech Day ARM has run, most in person but Covid made some virtual. They were all deeply technical but the last few years had marketing fluff intrude in ever more prominent fashion. For those who aren’t in the press, marketing BS is part of every presentation, more or less crushing depending on lots of factors. Most people hold their noses and play games on their phones while waiting for the boring stuff to end and the useful info to start. A prime example of this was the Intel server briefings where long time attendees would joke about the ‘pyramid slide’ before the briefs began. Jokes about last years slide with everything moved down a notch and a new top bit were never proven wrong. As you will see, ARM does this too.

In spite of this, ARM was the gold standard for technical briefs since day one, with emphasis on was. This year the ‘Tech’ day was literally devoid of tech, it was painful, insipid, useless, and insulting to the intelligence of attendees. SemiAccurate knows the people who made the content and they are universally intelligent, competent, and laden with other superlatives. This time around things were so bad that they either all suffered closed head wounds of a horrifically traumatic nature at once or they are hiding something. Also note that the people who delivered the message, and the ones who tend to wrongly get the blame for fiascos like ‘Tech’ Day, are just doing what they are told. They did their jobs quite well this time, the messages from above were the problem.

The next bit of Tech Journalist 101 is that when you have a briefing, there are tons of things that are not clear, this is a universal fact. “Did you mean X connects to Y directly or does it do so through Z, the diagram was a little unclear”. You walk up to the presenter during a break and ask about it. 30 seconds later you usually have clarity, rarely a second question drives things closer to the minute mark. Often you get even more insight or go down rabbit holes that educate and inform.

If the briefing is virtual there is usually a small window for Q&A afterwards that may allow a question or two from some attendees. That leaves anyone seeking real answers to write PR with a question. They forward the questions to the right people who then send the answers back to PR, and they usually send it back. Best case this cycle takes 2 days but quite often you get a response back that half answers the question or misses the nuance of what you need to know. Rinse and repeat. It usually takes a work week to get a clear answer. If that answer leads to more questions, well you usually run out of time before the embargo lifts, it does take time to actually write the story you know.

Back to this years ARM ‘Tech’ Day, and yes it was virtual. While this didn’t help things, virtual isn’t always fatal if done right. This year it wasn’t. Why? The content or to the point lack thereof. As you may recall, Tech Day is for journalists who actually care about the details and are informed about the industry at a low level. Most have attended past Tech Days and have a good handle on how things were done in the past.

This year the content of ‘Tech’ Day was so bad that in retrospect, it wasn’t worth our time to attend. If you think about the preamble to the book, “Baby’s First Agentic AI Play Day”, you are aiming too high. Yes EVERY talk was a high level overview style insipid spewing abut agentic AI, something that literally everyone in the audience is well aware of. It wasn’t even a technical talk about agentic AI like AMD did a few weeks ago at ADF 2026, it was marketing BS from start to finish. It was BAD. How bad? First go back and look at the 2015 picture above, then cringe as you go on.

ARM C2-Ultra CPU block diagram

“Details” were provided

ARM CPU layout details

Clusters have tech… Maybe

ARM Cluster performance ladder

The ladder moves up every year

The three slides above are from the ‘Tech’ Day deck on ARM’s new CPU line, the C-2 Ultra. They are literally the only ‘architecture’ slides about the CPU, and the Pro core is mentioned only two slides, the Nano core only in one. That one is the ARM equivalent of the Intel pyramid slide, and yes many of the same jokes are told about it. In any case this is a tech talk to tech journalists about ARM’s CPU architecture. The architecture literally wasn’t mentioned. Really. It was a joke.

Instead of pipelines, interconnects, and you know, tech, we got marketing BS on a level that just should not have been done. It was an embarrassment of riches without the riches. I can’t say this enough time, they literally did not do the very thing they were there to do, and what the deck called “CPU presentation” promised.

ARM C2-Ultra makes AI better

Feel the knowledge flow

So what did they do? They talked about agentic AI, well baby talked really. The slide above is the final of *5* slides that walk you through the high level of how agentic AI works, something everyone on the briefing knows in far greater detail than what was presented, not that there was any detail presented. The talk was about how ARM’s new CPUs when used as a lifestyle choice would make agentic AI better for you. We hope you never get to actually hear it, it isn’t worth the loss of brain cells.

ARM C2-Ultra is faster

Oh wow, faster is faster!

There were several slides presented with graphs clearly showing that the new cores were faster than the old cores, and when the new features were recompiled for, it was faster still! *GASP* Who knew? A new CPUs with new instructions outperforms the old one? Hold the freaking presses, we have a paradigm change here! Please note that other than the names, those new features were not detailed. There was nothing about how they actually did what they did. What was improved? How was it laid out? Details? Nope, just more pablum about agentic AI making you happy.

ARM C2-Ultra features

Yup it can get dumber

See what we mean? Larger execution windows are better, no sh*t. Smart speculation is more efficient? Really? Wow, that sounds interesting, when we get to the details there should be a lot of good here. But they didn’t get to the details, just circled back to agentic AI as a lifestyle choice over the next few slides. It was painful. There were more graphs that all said two things. First was that the CPU performance gain was roughly what TSMC has been promising from their process improvements, maybe a hair more. So this CPU line isn’t actually better at it’s heart unless you use the new features and recompile. Second is that new features add performance, consider me shocked!

Those new features were occasionally mentioned but there were literally no details. Nothing. ARM said there were big architectural changes but went out of their way to not talk about them, only how agentic AI was better with them, sans details. The Ultra core was talked about but the other two lines were not mentioned other than the slides we showed above. All new? Who knows? But they make agentic AI so much better. Somehow. Maybe. Who cares.

By now you are probably thinking this ‘Tech’ briefing was bad, you might even understand why we said it was possibly the worst briefing SemiAccurate has attended in 20+ years. That said the CPU briefing was by far the more technical and on-topic part of the day, the GPU briefing was worse, far worse. But it was about agentic AI, and AI is now infusing graphics!

ARM Mali G2-Ultra NX biggest change

Big changes, no details

If you look at the slide above, you may notice the bit about the new execution engine touted as the, “Biggest ISA re-architecture in 7 generations”! Wow, that is a big deal. I mean the last time they made a big architecture change you got slides like the ones below, and many many more.

ARM Mali Valhall goals

To dream of better (Tech) days

ARM 2019 Mali G77 arch

This is what Tech Days should be

ARM 2019 Mali G77 vs G76

This is what a real change looks like

Properly braced for technical details, the talk turned to, wait for it, AI-native graphics. Lots of partner testimonials, lots of demos, and lots of talk about how much faster it is to use AI rendering techniques that Nvidia and AMD have been using for 3+ years now. That went on for 14 slides, all talking about fluff and idiocy that everyone in the audience knew inside out an backwards. Then a breakthrough, real tech about the Neural Accelerator ARM keeps saying will help AI give you a warm hug. FINALLY.

ARM Mail G2-Ultra Neural Engine

Neural Accelerator fluffy overview

Just kidding. Once again it tells you what it does but not HOW it does. How is the reason that ARM does Tech Days and this time around doing the job correctly seems like a firing offense at the company. Lots of superlatives and lots of blindingly obvious claims but no details. None. Zip. Nada. It was pathetic.

ARM G2-Ultra execution engine

Mali G2-Ultra NX Execution Engine

Finally ARM talked about the new execution engine. Well not really, but they are quite proud that it is the biggest ISA upgrade in seven generations, you can see that on the slide. Other than three, two really, high level bullet points that any vaguely tech-curious person probably understands, that is it. Compare and contrast this to the slides from 2019 for a greater understanding of why we mock what ARM inflicted on us.

To be fair the following two slides could be considered tech if you are charitable and care about things at a very high level. One compared the G1-Ultra to the G2-Ultra for geometry elimination during raytracing, but this is what they claimed last year. Since there were no details, we have to assume it is just marketing BS trying to seem new. Again. Same with the next slide on hardware opacity micromaps, they literally said the same thing last year only this year it is different for reasons they won’t go in to. Pathetic.

As with the CPU IP lines, the GPUs come in multiple levels with distinct feature sets. This year the G2-Ultra NX replaces the G1-Ultra no NX, at least we think it does. The other levels were not mentioned at all, do they still exist? The layout of the G2 GPU changes from the G1… nope that part wasn’t mentioned either. The feature differences between the levels of GPUs as you move down the stack are… nope not mentioned either. How about 3rd grade level block diagrams? Absent. The GPU briefing was a joke but ARM did tell you about how graphics are now AI-native. Again. And again. And again. No details but it is AI-native, did we mention that? There are currently heavily sanctioned countries in the world that won’t do this kind of thing in their worst prisons, but it is AI-native.

The GPU section of the briefing was both long and useless. ARM literally did nothing that they were supposed to do at ‘Tech’ day and gave us all another long, lowbrow, boring talk about AI. They said there were block diagrams and levels but SemiAccurate could not find them in any of the materials provided. If they were supposed to be sent, they were not. The biggest architecture change in seven generations should have been something worth talking about, any guesses why they didn’t? More on that topic later.

ARM CSS for Mobile 2

Finally (barely) more details!

Then ARM went on to talk about the platform, CSS for Mobile 2. Please note this is another name change for no reason, last year ARM made a big deal about calling this Lumex. The assembled press at the 2025 tech day mocked ARM for this but ARM was adamant that this change would herald a new direction in branding for the company. Gone was the old CSSx and in with Lumex for a brighter future. Didn’t last long. I think this is the fourth name change in five years but I am honestly to annoyed to look it up. In any case ARM branding talk is slightly deeper than their tech talk which is to say at best it exists at a cursory level.

ARM SI L2

Interconnect ‘details’

The interconnect side had some more ‘details’ if you consider details to be a high level diagram surrounded by blindingly obvious features that give you a warm hug because they make agentic AI better. ARM won’t say how it does it, but since it is agentic AI you should be happy. I think we have used the term pathetic too many times already and my thesaurus is a bit out of reach so you think of an appropriate word.

This is all followed by talk about AI software offerings and developer tools that we couldn’t care less about. No details but it will make your life better and help you make better agentic AI for consumers. That was about it for the ‘Tech’ Day, they took a good thing and ruined it for no gain, just to hide something. Again more on that in a bit but first the next briefing ARM did that ties in to ‘Tech’ Day, and not just because it was painfully awful and insipid.

ARM CSS for Mobile 2

Fluffy but stupid AI drivel

ARM’s next briefing started out with, wait for it, agentic AI! Yay! Same idiotic low-level crap as before, just more. Who knew AI could be used in edge, cloud, and physical arenas? Everyone you say? Sure but ARM can take longer to say it, and dumb it down far more than you probably give them credit for. Keep reading and you will see I am not kidding.

ARM CSS for Mobile 2

More details but less than the bare minimum

Here we go again, a high level fluff briefing and ARM is giving more details than they did at ‘Tech’ Day. Really? Yes. Well, not MUCH more details, but at least they called out some units without explanations which is better than the ‘details’ of ‘Tech’ Day. Then they rehashed the same stuff as the last brief, some of the slides were directly pulled from that deck, and they told us how ARM makes agentic AI better and how we are in the era of AI-native graphics. No details, but write good things about us because agentic AI. It was that bad.

The next chunk was about physical AI and it centered on market size. No details but ARM claims to be selling lots of chips to a huge and growing market. Take their word for it, buy stock, and don’t ask questions. They claim to have sold two billion units into this market in 2025 but won’t mention a single one, what ARM adds, or how it works. But AI. Come on, this is sad.

ARM 500 million

Big numbers, don’t actually think about it though

Then we come to the datacenter side, a market so successful they once again won’t talk about details. ARM is supposedly selling billions of cores, note the change in metrics from above, into the datacenter market, but no mention of details.

ARM CPU creates intelligence

Derp derp derp, they are this bad

But it makes AI better, see above for how. No don’t, see above for marketing BS that is blindingly obvious but goes on for several more slides. For details they mention that they have two server CPU lines, Neoverse N and V, no mention of the third line but as you probably know by now, that is the new way of messaging at ARM. The details, no we can’t even joke about it by now, you know the drill. At least they mentioned the AGI CPU they are now competing against their customers with, but again they won’t disclose details.

ARM Neoverse CSS N4

Server ‘details’ lack the details part

Above is the most detail that ARM has given out about AGI, there was a talk at Hot Chips that was higher level than previous Tech Days, no scare quotes because it was in the past, something SemiAccurate hopes to cover at a later date. And that is about all ARM is willing to say about server, no datacenter tech days any more so live with stupid fluff, it will make AI better you know.

This all brings us back to the real point of the story. As we mentioned near the beginning, the people we know at ARM are competent and smart. They know what they are doing and why they are messaging things the way they do. We are not talking about the front facing PR people who will wrongly get the blame for this fiasco, we mean the people who architected the talk. They either all got stupid at once or there is something else going on.

As SemiAccurate said earlier, this sh*tshow was not just staggeringly awful, it was virtual. It was followed by a Q&A session that, at least for the sessions we were in, didn’t run the full time. Why? There was nothing to ask questions about. ARM personnel said to ask questions during this Q&A but, well, how can you ask questions about something you don’t know?

The largest architectural change in seven generation on the GPU is a good example of this. What do you ask? What are the changes is the obvious one but ARM would not talk about that, other than to say it made agentic AI better, warm hugs all around! Seriously though without block diagrams, without knowing the basics of the CPU/GPU layouts, what can you ask? You don’t even know the names of the new units they added so how do you ask about them? At best you can start to go down the path of teasing out the basics but there was one hour shared for multiple people, good luck getting anything useful.

For the first time in our memory the people on the call didn’t have enough questions to fill the allotted time. If you have even done this kind of work, you know how stunning that statement is. So SemiAccurate for one spent some of that time bitching about how awful the ‘Tech’ Day was bur to no avail. To this day we have no idea what the 2027 ARM offering actually are, we only know a subset of the level names on some of the products. Tech Day my ass.

So circling back to the point of why, we will rule out mass closed head wounds that degrade logic and communication skills to the point of public drooling although that would probably have resulted in a more useful presentation. Instead we will bring up the topic that we were hoping we could get shot down at Tech Day but were denied the chance.

One of the best uses of time at events such as Tech Day is to talk to the presenters and ask them about things you have been hearing. “X is saying Y about your Z, that isn’t true is it?” This often leads to a conversation about how X is bending the truth about something or even outright lying. Occasionally you get them to say X isn’t all that great in one aspect but here are the reasons it isn’t a big deal in practice. In short it is a great way to shoot down rumors and almost always greatly benefits the company.

This time around we were hoping to ask ARM about a few topics, starting with their CPU architecture and why it is in fourth place in a three company race. We couldn’t but the slide above where they show that their performance without the added accelerators is roughly in line with TSMC process improvements says it all. Why doesn’t ARM talk details? Because they know what the details will show, and they really don’t want that out in the public. If you look at the performance of Apple and Qualcomm silicon over the last few generation, it does not shed a good light on ARM’s architectural progress. If I were laying out the Tech Day talks, I would not have made the woes this obvious.

Go back and look at the slide with 1.5+ Billion Neoverse cores shipped into datacenters globally, that is a big number! Better yet 500 million of those were shipped in the last 9 months, so the growth curve looks to be pretty steep. Unless you stop and think about it, then things are a little less rosy.

What do we mean? Take the 500m number and think about that. They quoted cores not devices for a reason. How many cores does the average Neoverse device shipped into datacenters have? 64? 128? 192? AGI has 136 per die, Microsoft sports 128 if memory serves, and Amazon Graviton 5 has 192, but if you count things like smartNICs that average goes down. Amazon is the only one of these shipping in any sort of volume, it you take them out of the equation that 500m number goes down a bit, our educated guess is to around 7. That is seven units, not 7m. Microsoft made a token number of their SoCs to use internally, Ampere does not use Neoverse cores nor does Nvidia (officially), and there really isn’t anyone else.

Seriously though, those numbers are not good. They are big but if you look at the x86 server market, it is 30++m SOCKETS a year but each of those sockets average vastly more cores per socket than ARM. This is the long way of saying those server numbers ARM is touting look good to the uneducated who don’t think about what they are seeing but pale in comparison to the competition.

This is another biggie that ARM is trying to hide, their server offerings are not doing well. Similarly their PC efforts are dead in the water, 12 years on and there isn’t a functional ARM PC on the market. There are literally three of them by my ankle as I write this and none rise to the level of working right. Any time ARM steps out of their traditional markets, embedded and mobile phones, they fall flat. Why? Their architecture is way behind the competition and the gap increases with each generation.

That is why ARM won’t talk about tech at ‘Tech’ Day. They know what they have and they know it isn’t good. If you don’t have it, distract. If that doesn’t work, change the subject or talk about something adjacent to your problematic parts. In this case instead of a tech day, we were treated to hour upon hour of high level pathic babble about agentic AI and how AI infused graphics would make everything better because AI. It was honestly the worse briefing we can recall in 20+ years, and it was bad for cause. Too bad it didn’t work to distract.S|A

The following two tabs change content below.

Charlie Demerjian

Roving engine of chaos and snide remarks at SemiAccurate
Charlie Demerjian is the founder of Stone Arch Networking Services and SemiAccurate.com. SemiAccurate.com is a technology news site; addressing hardware design, software selection, customization, securing and maintenance, with over one million views per month. He is a technologist and analyst specializing in semiconductors, system and network architecture. As head writer of SemiAccurate.com, he regularly advises writers, analysts, and industry executives on technical matters and long lead industry trends. Charlie is also available through Guidepoint and Mosaic. FullyAccurate