ExpertsFM20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · September 19, 2026
5 Selected Lessons:
case · 3 framework · 1 heuristic · 1
Compute-bound training vs memory-bound inference
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron (September 19, 2026) · at 4:14 ▸ · at 7:07 ▸
When this applies
when diagnosing whether a workload's bottleneck is raw processing power or data movement/retrieval
The takeaway
Separate the 'produce' and 'consume' phases of any system and ask whether the real constraint is throughput capacity or access latency to existing state.
Why it works
Training can parallelize over a known corpus so throughput scales with flops, but generation is sequential and must read the full parameter set for every single output step, making it bound by how fast you can move data, not by how fast you can compute.
One level up
The resource you optimized for in one phase of a system's lifecycle is often the wrong resource to optimize in the next phase.
Where it stops holding
This stops working when a workload actually can be parallelized ahead of time (e.g., batch processing with full foresight), in which case it behaves like a compute-bound problem regardless of being 'inference'.
Detail
procedure:
- Identify whether the workload has access to the full input set upfront (training) or must generate outputs sequentially without foresight (inference).
- If parallelizable, optimize for flops/compute throughput.
- If sequential/auto-regressive, optimize for memory bandwidth and capacity instead.
rationale: Every generated token requires reading the full weight set; this read cannot be skipped or batched ahead of time the way training data can.
In their words
But when you're inferring, because that's actually generative, you don't know what the token is five words down the line. And so you have to generate each and every one auto-aggressively or in order without foresight. And so that becomes a hugely memory-bound problem that can't just be massively parallelized, like training.
The memory wall: compute scales faster than memory
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron (September 19, 2026) · at 7:07 ▸
When this applies
when any system's two dependent subcomponents have historically improved at wildly different rates, creating a growing bottleneck
The takeaway
When two coupled technologies have different innovation curves, the slower one eventually becomes the system's binding constraint even if it was never the historical bottleneck.
Why it works
SRAM memory cells have not shrunk with Moore's Law the way logic transistors have, because the underlying architecture is decades old and there was no strong market incentive to redesign it until AI workloads demanded it.
One level up
Track the growth-rate gap between coupled subsystems, not just their absolute levels, to predict where future constraints will emerge.
Where it stops holding
This stops working when new memory architectures (not just process shrinks) emerge that break the historical SRAM scaling pattern, or when workloads shift back to being compute-bound.
Detail
rule: Expect memory bandwidth to lag flops growth by roughly an order of magnitude over a decade in GPU architectures.
threshold: 120x improvement in flops vs only 17x improvement in memory bandwidth, 2014–2024
rationale: SRAM's six-transistor cell design has not scaled architecturally in 30-40 years, while logic transistor density kept improving with Moore's Law.
In their words
So roughly speaking, between 2014, just very early innings of the new AI era, till 2024, you had about 120x improvement in the flops of GPUs. So a single NVIDIA GPU had about 120-fold improvement in flops.
Debunked data-center resource myths vs. actual self-generation economics
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron (September 19, 2026) · at 0:00 ▸ · at 19:51 ▸
When this applies
when public opposition to infrastructure is based on viral but unverified comparisons rather than actual resource accounting
The takeaway
When evaluating infrastructure opposition, separate the actual resource flow and allocation mechanics from viral anecdotal comparisons that may not hold up to scrutiny.
Why it works
Modern data centers use closed-loop liquid cooling and typically bring their own generation capacity that covers their own use and beyond, so they are not drawing from or raising prices on residential allocations, even though viral claims suggest otherwise.
One level up
Strategic adversaries can exploit asymmetric openness — one side's freedom to spread unverified claims becomes a structural disadvantage against an opponent unconstrained by the same discourse norms.
Where it stops holding
This stops working when a given data center genuinely does draw from already-allocated grid capacity rather than bringing new generation, in which case the resource competition concern is legitimate.
Detail
situation: Bipartisan anti-data-center sentiment has grown based on claims about water and electricity usage.
decision: Sohmers argues these claims are largely false — a single In-N-Out uses more water than the largest US data centers, and new data centers bring their own generation capacity rather than drawing from residential allocations.
outcome: He characterizes the anti-data-center narrative as strategically beneficial to China, which is expanding dirty generation and bulldozing communities without the same domestic political friction.
lesson: Open societies' freedom to spread unverified claims can become a geopolitical liability if it slows legitimate infrastructure buildout relative to less constrained rivals.
In their words
The thing from water usage and like the amount of power they consume, et cetera, like so much of the early information that went out by unsophisticated writers or things that are just patently false. Like, you know, a single in-and-out uses, you know, more, more water than, you know, the largest data centers in the United States.
GPT-6 Astra automates a 2-3 week chip design flow in ~50 hours
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron (September 19, 2026) · at 47:44 ▸ · at 50:25 ▸
When this applies
when judging whether an AI tool has crossed a real capability threshold, by testing it against a task that has consistently failed with prior tools
The takeaway
To evaluate whether a new AI capability is a real leap, test it on a task that is hard specifically because of poor documentation/tooling, not just raw difficulty — success there is a stronger signal than benchmark improvement.
Why it works
The task (chip design from spec to tapeout) had poor public documentation and outdated tooling that previous models couldn't handle even partially; succeeding fully end-to-end on such an out-of-distribution, under-documented task signals genuine new capability rather than incremental benchmark gains.
One level up
The best way to detect a genuine capability jump is a task selected for being unsolvable by prior generations, not a task from a standard benchmark suite.
Where it stops holding
This stops working as a universal signal once models are specifically trained/fine-tuned on this exact class of task (the test becomes 'teachable to the test'), reducing its diagnostic value for other out-of-distribution capabilities.
Detail
situation: Every previous model had failed to take a simple logic design problem (an encryption block) through the full specification-to-tapeout chip design flow, partly because EDA tooling from the 90s/2000s is poorly documented.
decision: Sohmers tested GPT-6 Astra on the same task, leveraging its computer-use and scripting abilities.
outcome: Astra completed the full flow and met timing at over a gigahertz in just over 50 hours, compared to an estimated 2-3 weeks for a competent human new to the task.
lesson: A model's ability to solve a previously-intractable, poorly-documented real-world task is a stronger signal of genuine capability jump than incremental benchmark score improvements.
In their words
GPT-6 with both combination of computer use and just an ungodly amazing scripting ability has been able to take this KechAck block and implement it with the TSMC and the three PDKs and take that all the way to GDS and do that in like a little over like 50-something hours and meet timing at you know over a gigahertz.
Chinese labs innovate algorithmically around hardware export control constraints
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron (September 19, 2026) · at 53:57 ▸ · at 57:12 ▸
When this applies
when a resource constraint is imposed on a competitor, expect them to innovate around the constraint algorithmically rather than simply falling behind
The takeaway
A hard resource constraint imposed on a capable competitor often produces algorithmic innovation that substitutes the constrained resource for one that is more available, rather than simply reducing their output.
Why it works
Denied access to the highest memory-capacity and flop chips, Chinese labs developed techniques (multi-head latent attention, gated DeltaNet-style mechanisms) that trade additional compute for drastically reduced memory/KV-cache requirements, achieving comparable capability through a different resource tradeoff.
One level up
Constraints are generative: forcing a capable actor away from an abundant resource pushes them to discover efficiency gains that the unconstrained competitor never had reason to pursue.
Where it stops holding
This stops working when the constrained actor lacks the technical talent or research capacity to find the algorithmic substitution, in which case the constraint simply reduces their output as intended.
Detail
situation: Export controls restricted Chinese labs' access to chips with the highest memory capacity and flops.
decision: Chinese labs developed multi-head latent attention (MLA) and related techniques (e.g., Gated DeltaNet) that spend more flops to drastically shrink KV cache size and attention compute time.
outcome: DeepSeek v3 achieved major KV cache size reductions; Gated DeltaNet derivatives cut attention computation time by about 75%, though Sohmers notes these come at some cost to model capability and major US labs reportedly have not adopted MLA.
lesson: Export controls on a specific resource can backfire by accelerating the constrained competitor's innovation in substituting away from that resource.
In their words
The algorithmic advancements that have happened over the past year have been very, very interesting in terms of being able to further reduce the amount of storage and compute necessary for that context with linear and sparse attention mechanisms. And those have really, really been innovated by the Chinese model labs.