Semianalysis AI Value Capture – The Shift To Model Labs Anthropic is now making $44 billion per year run rate and this is heading to $100 billion per year by the end of 2026.
As of today, Memory SOCAMM contract pricing paid by Nvidia at ~$8/GB in 1Q26, a sharp step-up from 4Q25 to 1Q26. Predicted exit ’26 pricing for SOCAMM could exceed $13/GB, which is roughly in line with mobile DRAM pricing expected by the end of this year; accordingly, we view ~$10/GB as a reasonable assumption for Nvidia’s SOCAMM cost.
Vera Rubin VR NVL72: V for Value – Rubin delivers a step jump in performance per TCO.
ROI is going more strongly to AI model makers like Anthropic and to the memory companies.
Capex Per Watt Trends from GB300 to VR NVL72
For GB300, DRAM was bundled into the board and marked up at ~75% gross margin, making the margin charged on the memory on the board consistent with what is implicitly priced for the Blackwell systems.
For Rubin, we initially assumed the same dynamic, with the understanding that Nvidia would target an overall system Gross Margin in the mid-70s. As such, our initial Bill of Material (BoM) modeling applied a consistent margin throughout the entire Strata board leaving SOCAMM margin at the same mid 70s margin.
Hyperscaler networking cost advantages over Neocloud only becomes a 10% increase in all-in capital cost for a full rack-scale server.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Meanwhile Grok 4.3 scores 34 spot on LLM arena with rating 1456, while 4.2 versions have good ranking.
I find if odd. I guess it is a miss for some reason.
I mean Grok has good real time data(multi agent approach), while some other models are using old data and think they are way in the past. That is good and some versions like 4.2 are good in general.
When will supply meet demand? Where’s the plateau?
How many fab plants are under construction?
I know full well this will come off as whining (which is exactly what it is), but there very much needs to be a way to keep consumer RAM, disk, GPU, and CPU prices from scaling upward with LLM demand. The consumer market should not be pegged to A.I. demand. I know some of it is related to materials, but it’s honestly ridiculous. But if this is just something that *feels* like it’s happening as opposed to reality, please correct me (seriously).
People who enjoy doing their own local language model work, fine-tuning, etc, as well as people who like to build their own high-end gaming PCs have to pay enough for RAM already.
32 GB DDR5 6400 and 2 TB NVMe should absolutely not cost a massive sum of money. If that sounds like a comical statement, it’s not meant to be. But that said, I do remember it seemed absurdly expensive to add 4 megabytes of RAM to our old family PC when I was a little kid.
Probably shouting into the void, but either come up with new A.I. architecture that uses organic neuromorphic computing with materials like smart polymers (which seems wild and exciting), or come up with new consumer hardware so hobbyists don’t get thrown under the price bus.
“Yeah, I know, nothing can ever please me.”
–Me
It was pretty expensive to add 4K RAM to a computer back in ancient times.
So true, good point. I suppose it seems that, as technology becomes more advanced and hopefully more efficient, it should become cheaper.
But that mindset doesn’t take into account potentially dwindling supplies of raw materials and their costs (which relate to a host of things like geopolitics), nor those supplies being diverted for adjacent use (in this case, LLMs).
My hope for A.I. in general, and even for consumer computing, is that new technologies are created that are less reliant on rare materials (or materials that could become rare in the near future).
The agentic code tools have demonstrated being far more useful than the funny chatbots, the images, video and other generators.
Almost every company will be turning them into their information backend soon.
It’s not a fad, but a new computing paradigm.