What becomes free and abundant stops being where the money is and becomes less important. Tap water is essential and it is not an industry. If Jason is right that Deepseek and Other Open Source AI are indistinguishable for 95% of queries. The consequence is not that corporations buy hardware and run open source. 95% of what we are doing becomes worth roughly nothing per token, and all the remaining value concentrates in the 5% where reliability, liability, and the last few points of per-step accuracy decide outcomes.
It’s happening
Default will be local model by next year for most corporate workers https://t.co/wcTL8RFKLd
— @jason (@Jason) July 31, 2026
All companies would the same problem that AI cloud providers have. Capture enough value for return on the hardware spending before it is deflated. There is a 10X annual deflation in the tokens.
The deflation trap and the agent-chain panel are really the same wall viewed twice. Commoditization is what makes local capex unrecoverable, and it’s simultaneously what pushes the money into exactly the regime where local can’t compete. His mechanism is right; it just doesn’t terminate where he says it does.
What is Likely to Happen
1. Open weights on rented cloud / neoclouds (~50–55%)
Model quality and price pressure play out exactly as Jason describes. MIT-licensed (or similar) weights dominate token volume and cap closed pricing. Almost none of the volume runs on customer-owned hardware. This is already the base case and continues.
2. Cascade / router hybrid (~20–25%)
Small local or ultra-cheap cloud models handle classification, simple generation, redaction, autocomplete, and the easy majority. Frontier (open or closed) handles escalations. Local tokens can grow 10–50× while local spend share stays small. This absorbs 100× local tokens without requiring primary local dominance or solving the memory/fit problem for frontier models.
3. Local arrives as invisible OS feature (~10–15%) — Apple, Google, Microsoft, etc., ship competent small-to-mid models inside the OS / on-device for private, low-latency, offline tasks. Users never switch. It just works for the 80% easy cases. Currently blocked by falling memory content per device and the quality gap on anything agentic. Becomes more plausible if architecture efficiency delivers 5–10× memory-per-useful-token gains.
4. Sovereign / regulated / air-gapped islands (~5–8%) — Real, growing, and defensible for specific sectors. Structurally capped at low single-digit percent of total tokens and higher (but still limited) share of high-value regulated spend.
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.
The money is in the hardware
Right now many entrepreneurs produce largely digital products and do digital work
All of that will be free in 2-3 years time
Distributed manufactuting may be the way out but even much of that will be done by AI
CAD/CAM design work will be free
Making the product will be outsourced to distributed fabs
An AI will be capable of doing this as will a human
While it’d be a different direction and so would need a lot of work, there might be another path, that keeps company data local, uses a small local Ai to generate simulated data similar to the real data the company wants to keep private, and then gives a cloud Ai that data and the task of writing code that can use a small local LLM (as well as “normal” data processing code) to do whatever processing/thinking the company wants done on their real data.
I tend to ageee but the current issue is the hardware required to run it locally is in deep shortage and will not be resolved any time soon, try to get a Mac to run it , several month delays
I can see the value of most A.I. results plummeting to almost zero already in my work for two NGOs. I’ll be asked to do some report, or just do it myself, prompting with my knowledge gained over nearly 2 decades. A report will be produced in under an hour, even with a series of prompts to refine the conclusions. The report will be either acted upon or, more likely, discussed for a while then shelved. The next time there’s a similar request, no one digs around the library or even their own memory for previous reports, it’s just done again, maybe with the latest nuance in the request. The cycle repeats.
This is the result for anything that is free – which is what 95% of A.I. is now in order to gather users in the hyper-competitive world of LLM A.I. If it’s free to use, it has no value. It doesn’t matter if it’s tap water or sophisticated LLMs given freely to the public, even though they are far from free to run; the average ChatGPT request burns 10X the energy alone as the average Google search engine request and even THAT baseline request is not free, it’s paid for by ads. But ads don’t produce enough revenue, or are simply not used (or wanted) by users, to pay for A.I. results. Even backdoor subscriptions, like xAI’s X premium subscription model, does not pay enough for Grok among heavy users of it (I use it nearly every day from my premium account on X).
Models will get cheaper, but never THAT cheap. Also, China is subsidizing their A.I. industry, as they do for any industry they eventually come to dominate. And China pays for that by artificially devaluing their currency, which does not float freely on the world Forex market. They’re masters at that and it barely gets a mention anymore.
Great point about China subsidizing A.I. They should be iffy about weakening the dollar. Because if they end up having to free float their own currency, their current way of subsidizing business goes out the window. They’d fall further behind. But then again, my understanding of how all that macroeconomics works is wobbly.
The more efficient models become, and the more efficient hardware to run the efficient, dense models becomes, and the cheaper local fine-tuning becomes, people who can afford the hardware to run local will buy the hardware.
A scenario in which hardware and software is purposely made unavailable in order to prevent people from locally running fantastic models with no guardrails is a pipe dream.
Open-source will always be invincible.