Get ready for another chance to hear from the leaders behind Atombeam.

This Friday, we’ll be going over what’s new and exciting – and it seems like every week there is something new and exciting.

We’ll answer live questions from the audience, and you never know who will pop in. So get your tech prepared:

• Download the app and log in (Apple App Store | Google Play Store)
• Follow Atombeam and sign up for push notifications
• Allow notifications in your settings

We’re scheduled to go live Friday, August 7, at 10am PT / 1pm ET.  Mark your calendars, and we’ll let you know when we’re kicking off.

[View Offering Details]

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.

‍

Hi Everyone,

Something happened in the AI market this spring that I think matters more to Atombeam than any benchmark release, and I want to walk you through it. It has been covered in the trade press as a US-versus-China story. I read it differently, and the difference is the point of this note.

What happened

OpenRouter is the largest aggregation platform for AI models. Developers route their API calls through it to reach hundreds of models from a single interface, and because the platform publishes detailed usage data, it serves as a barometer of what developers actually run in production, as opposed to what they praise in benchmarks. In the week of February 9 to 15, Chinese models processed more tokens on the platform than American models for the first time: 4.12 trillion against 2.94 trillion. The gap has since widened sharply. By June, Chinese models were processing roughly 18 trillion tokens per week against about 5.5 trillion for US models among the platform's top models.

The cause is straightforward. Usage has shifted from chat to agents and coding. Programming went from about 11 percent of platform volume to more than half over the course of 2025, and agent workflows now generate the majority of output tokens. A single overnight agent run can invoke a model thousands of times, so at that intensity the price per million tokens becomes the dominant operating cost. The Chinese labs price accordingly: MiniMax's flagship lists at roughly $0.30 per million input tokens and $1.10 per million output tokens, against $5 and $25 for the premium American model it most often displaces. Developers did the arithmetic and moved.

Why we think this matters to Atombeam

First, the market now votes on cost per task. We have been arguing for some time that the contest in AI is shifting from raw capability to cost per outcome. We believe this spring was the largest natural experiment yet run on that question. Developers left benchmark leaders for whatever they could afford to run at scale, and cost won.

Second, there are two ways to win a cost contest, and only one of them is a price war. The Chinese labs cut the price of computation, supported by cheaper energy and efficient serving. That is a price war among systems that share the same underlying architecture, and a price war among identical architectures races to a floor set by electricity costs. PCM takes the other path: it is designed to cut the quantity of computation.* A transformer re-derives its knowledge on every query, and every million-token agent run is paying that recomputation tax. Discounting the tax still leaves you paying it. Our prototype measured roughly 100x per-query compute efficiency, and 840x on learned subjects against a small language model. Those are prototype numbers on prototype workloads and we treat them as such, but the mechanism behind them, persistent knowledge that is not rebuilt per query, does not depend on winning an energy-price war.

Third, we believe the market is splitting into two tiers, and the split may be our opening. In the commodity tier, the cheapest token wins, and today that means a Chinese model. In the premium tier, enterprises in regulated industries pay multiples of commodity pricing for safety, compliance, and trust. Here is the structural fact underneath the headlines: the premium buyers cannot take the Chinese discount. Every major Chinese model operates under China's National Intelligence Law, which obligates its maker to cooperate with government intelligence requests regardless of where the servers sit, and DeepSeek is already banned on government devices in the United States, Taiwan, South Korea, Italy, and Australia. So the customers we are targeting first, defense and regulated enterprise, sit trapped on the premium price curve while watching the commodity curve fall away beneath them. Our Private AI pitch is aimed at exactly that gap: the properties they pay the premium for (data behind the firewall, no sovereignty exposure, auditable decisions, and refusal instead of fabrication) delivered at economics designed to undercut the commodity tier. We believe nobody currently offers both, and this spring made the gap wider.

Fourth, the labs' own margins now depend on architecture. If token prices commoditize toward a floor set by Chinese energy costs, an American lab cannot defend its margins with scale, because scale is what its competitors are discounting. Margin defense has to come from a different cost structure. That strengthens the argument we made in the Wider Picture note: the frontier labs are among the parties with the most to gain from an architecture that changes the cost structure rather than the price sheet, whether as licensees or as something more. And the industry's own leading voice on scale has conceded the underlying point. Ilya Sutskever, the OpenAI co-founder most identified with the scaling thesis, said in a November 2025 interview that he does not believe another 100x of scale would transform capability, and that “it's back to the age of research again, just with big computers.” When the architect of the scaling era says the next gains come from research and architecture, we believe the value of a genuinely different architecture rises.

What could cut against this

OpenRouter is a slice of the market, and it skews toward cost-sensitive developers. The most-cited share figures measure the top 10 models in single weeks; a broader OpenRouter study across 100 trillion tokens still put Western proprietary models at roughly 70 percent of total global API share. We think the trend is real but the headline ratios overstate it.

Enterprise prices could fall faster than we assume. Our savings framing rests on enterprise AI spend staying on the premium curve. For regulated buyers we believe that holds for the structural reasons above, and consumption is growing faster than prices are falling (Goldman Sachs projects a 24-fold increase in agent-driven token consumption by 2030), but it is an assumption and we treat it as one.

Our own numbers remain prototype numbers. The efficiency figures carry their qualifiers, the zero-hallucination result is scoped to our testing to date, and the expectation that PCM's cost falls with use is a design property we have not yet measured on the prototype. We will report measurements when we have them.

So my read is

The trade press calls this the flippening and covers it as a national contest. What actually happened is that the market repriced AI on cost per task, it did so in a matter of weeks, and in doing so it revealed a large customer segment that wants the low-cost curve and is legally unable to buy it. We believe that segment is our launch market. The Chinese labs proved the demand is real. The sovereignty wall proves the premium buyers cannot follow the discount. We believe PCM is built for exactly the space between.

As always, do the math yourself. Take an agent workload of a million tokens per task. Price it at $25 per million output tokens on the premium curve and $1.10 on the commodity curve. Then ask what a regulated buyer, who is barred from the second number, would pay for the first tier's trust properties at below the second tier's cost. That gap is the business we are aiming to build.

Charles

This update contains forward-looking statements, including statements regarding product development, expected performance, market conditions, and business strategy. These statements reflect our current beliefs and are subject to risks and uncertainties that could cause actual results to differ materially. Prototype results may not be indicative of production performance. Third-party market data is drawn from sources we believe reliable but have not independently audited; a companion source-notes document is available.

*The PCM technology is still in development, and there are substantial technical risks that could prevent us from achieving these efficiency gains at scale. Competitive advantages in technology are often temporary, and competitors may develop alternative approaches that match or exceed our efficiency claims. Market adoption of power-efficient AI is not guaranteed, and regulatory, technical, or economic factors could impact the viability of our approach. This technology has not been validated in large-scale commercial deployments, and significant engineering challenges remain before commercial release. Investors should consider this a high-risk, early-stage technology investment with uncertain outcomes.

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.

Hi Everyone,

I have been reading a few articles this week that, read together, seem to make a case for PCM in a new generation of processors that use neuromorphic computing.* Forbes ran a piece on where AI is headed over the next decade with a section on neuromorphic computing, processors modeled on the brain that in the author's words allow for continuous learning while using a fraction of the energy of modern GPUs. The National Academy of Engineering published a technical survey on low-power AI hardware. Nature ran an editorial on the same theme. And a Hebrew University group published a study in PNAS on why the human brain is so much more efficient than the machines we build to imitate it. I want to walk through what these say, because they describe the destination the whole industry is now pointing toward, and why we believe our Persistent Cognitive Machine, PCM, is potentially already most of the way there by a different road.

1. The paradigm the industry is trying to escape

The consensus problem is cost, a subject regular readers of these updates will recognize. The National Academy of Engineering survey lays it out. Two trends that drove computing costs down for fifty years are flattening: single-thread processor performance has been roughly flat since 2010, and energy per operation is approaching the physical floor of the technology. At the same time, AI models are outgrowing their hardware. The survey puts the mismatch in stark terms: large language model size has been growing at roughly 240 times every two years, while memory per GPU grows at about 2 times over the same period. That gap is what they call the memory wall, and it is why data centers scale by adding more GPUs rather than getting more from each one. It also seems to explain why memory chips are so pricey. Nature's editorial describes the same underlying constraint, the separation of memory and compute that forces energy-costly data movement. When the industry's most credible institutions are independently documenting the same ceiling, it quite possibly is real.

2. Why cheaper prices do not solve it

A reasonable person will say that costs in technology always come down, so this will sort itself out. It is a fair challenge, and the answer is in why costs came down. For fifty years the driver was Moore's law and its efficiency counterpart, Koomey's law, under which each chip generation did more per watt and per dollar almost automatically. Both are now flattening, as the National Academy of Engineering survey documents. Prices will still fall through competition, and cheap Chinese models are proof of that, but competition lowers the price of a given capability. It does not restore the exponential efficiency gains that came free from the hardware for half a century, and it does nothing about the memory wall. So costs will keep drifting down while the mechanism that made them collapse on their own runs out. We believe that is the moment when efficiency that comes from architecture, rather than from the next chip, becomes valuable.

3. What the industry says the answer looks like

Strip away the specifics and the proposed escape route has four properties. Systems that learn continuously instead of being frozen the day they ship. Systems that sip power instead of gulping it. Systems that run at the edge, close to where data is generated. And systems that adapt in real time. The Hebrew University study adds a deeper note underneath all this: it found that a single human cortical neuron rivals an entire deep network in processing power, because of its rich physical structure, and its authors argue that today's AI is inefficient precisely because it is built from, in their words, hyper-simplified, uniform units. The lesson they draw is that the sophistication of the building block matters more than the size of the pile. Hold that thought.

4. Where PCM sits, and where it is ahead

PCM is not neuromorphic hardware, and we do not claim it is. We believe it to be a software reasoning engine built on geometry rather than on simulated neurons or new silicon. But measure it against those four properties. On power and footprint, PCM can run on under 1MB of working memory, requires no GPU, and uses integer operations on ordinary processors. On the edge, it is designed to run on small devices, and 17 of our patent filings carry Edge or Mobile in the title, including three issued patents on mobile-optimized persistent cognitive architectures. On real-time adaptation, prototype response latency has measured 2.4ms. On efficiency, in prototype testing on learned subjects, PCM has run at roughly 840x the compute efficiency of a small transformer-based language model.

Continuous learning is where the comparison turns from parity into an edge, and it is worth being precise. The survey is candid that most neuromorphic processors today still depend on off-chip learning: they are trained on GPUs and then deployed only to run inference. Learning in the field, on the device, remains hard and rare even for the leading chips. PCM is designed to learn from use in deployment, which is the Persistent in its name. So on the one property the industry most wants and has had the most trouble delivering in hardware, we believe PCM already does it in software. And to the neuron study's point, PCM's efficiency comes from a more sophisticated building block, a geometric reasoning unit, rather than from piling up simplified ones. That is the same lesson the biologists drew, arrived at independently.

5. Why this matters for the investment case

We think that each of these pieces is a credible institution teaching the market that the future of AI is continuous learning at low power, outside the data center, from better building blocks rather than bigger piles. We believe that this is a description of the category PCM occupies, delivered by people who have never heard of us. We could not buy category education of this quality. So my read is that the neuromorphic wave and the science behind it may validate the destination, while the hardware path to it runs through new chips, new materials, and a developer ecosystem the survey's own authors place years out. And we believe PCM offers the same destination on a shorter road, in software, on hardware that already exists everywhere. The two are complementary, not rival: nothing stops PCM from running on neuromorphic chips if and when that ecosystem matures.

Caveats

PCM is pre-Alpha. The full Alpha is expected in September. The comparisons here are directional, not benchmarked: we have not run PCM against neuromorphic hardware, and our efficiency, latency, and memory figures come from our own prototype testing on learned subjects against a small transformer-based language model, not from independent evaluation. The neuron study and the hardware survey are third-party work we cite for the trend they establish, not as measurements of PCM.

One closing thought. The industry is preparing to spend years and enormous capital building new chips to get continuous learning at low power outside the data center. We believe we have those properties running in software today, in under 1MB, on processors you already own, with the one hardest property, learning in the field, seemingly already working. If we are right, the market's appetite is settled, and these pieces could potentially settle it.

Charles

*The PCM technology is still in development, and there are substantial technical risks that could prevent us from achieving these efficiency gains at scale. Competitive advantages in technology are often temporary, and competitors may develop alternative approaches that match or exceed our efficiency claims. Market adoption of power-efficient AI is not guaranteed, and regulatory, technical, or economic factors could impact the viability of our approach. This technology has not been validated in large-scale commercial deployments, and significant engineering challenges remain before commercial release. Investors should consider this a high-risk, early-stage technology investment with uncertain outcomes.

This update contains forward-looking statements, including statements about product development timelines, expected demonstrations, market direction, and anticipated performance. These statements are based on current expectations and beliefs and are subject to risks and uncertainties that could cause actual results to differ materially. Performance figures cited for PCM are from internal prototype testing under the stated conditions and may not be indicative of results in commercial deployment. Nothing in this update is an offer to sell securities.

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.