What E055 Got Right and Wrong About DeepSeek R1

What E055 Got Right and Wrong About the DeepSeek Shock

Venture Step E055 was right to distrust the idea that one cheap-model headline settled the economics of artificial intelligence.

It was wrong to let that skepticism become an unsupported theory about market manipulation. No source in the episode package substantiates fraud, coordination, or manipulation, so the revised story removes that insinuation.

The better question is narrower and more durable: What did DeepSeek actually report, which artifact did the number describe, what did it exclude, and what can that evidence say about hardware demand?

The February 2025 story collapsed several layers

DeepSeek R1 entered public discussion as a capable reasoning model, an open release, a low-cost training achievement, a challenge to U.S. export controls, and a threat to NVIDIA.

Those are different claims.

A model family is not a company. A checkpoint is not a hosted API. A final training run is not total research and development. Accelerator hours are not dollars. Lower compute per task is not automatically lower total hardware demand. A stock-price reaction is not proof that the underlying story was true or false.

E055 sensed that collapse. The episode waited for some of the first panic to settle and asked whether the low-dollar story was too neat.

That instinct was useful. The explanation that followed needs correction.

The famous number belonged to V3 training scope

DeepSeek's V3 technical report describes a 671-billion-total-parameter mixture-of-experts model with 37 billion parameters activated per token.

The authors report 2.664 million H800 GPU-hours for pretraining and about 0.1 million for subsequent stages, summarized as 2.788 million H800 GPU-hours for full training.

That is a defined compute statement. It is not a complete audited bill for DeepSeek, R1, or the company.

It does not automatically include every architecture experiment, failed run, earlier model, data pipeline, staff cost, hardware purchase, power system, facility, network, software effort, evaluation, service deployment, inference expense, or corporate overhead.

R1 also involved its own post-training process. DeepSeek's R1 paper describes R1-Zero, cold-start data, reinforcement learning, supervised fine-tuning, and distillation relationships that cannot be reduced to one V3 base-training number.

flowchart TD
    A["Reported V3 GPU-hours"] --> B["Defined training scope"]
    B --> C["Add conversion assumptions"]
    C --> D["Illustrative dollar scenario"]
    D --> E["Still excludes unreported company costs"]
    E --> F["Cannot become total R1 development bill"]

Multiplying GPU-hours by a public rental rate can produce a scenario. The result depends on the chosen rate, ownership, utilization, power, facilities, network, storage, labor, depreciation, region, and period. It is not a discovered invoice.

R1 was not one artifact

DeepSeek's official R1 repository distinguishes R1-Zero, R1, and six dense distilled checkpoints based on Qwen and Llama families.

The full R1 family builds from DeepSeek V3 architecture. The distilled models are smaller, separate artifacts. A third-party quantization or package creates another layer. The hosted API is a service with current names, terms, data handling, price, limits, and version behavior.

Calling every one of these "DeepSeek R1" makes evaluation and licensing unreliable.

The repository says its code and principal weights use the MIT License, while distilled models retain obligations from their base families. "Open source," "open weight," and "commercial use" should therefore be applied to the exact artifact and dependency chain.

The episode's manipulation theory had no evidence

E055 speculated that the release and market reaction could have been used for financial gain.

The episode did not provide trading records, communications, positions, coordination evidence, regulatory findings, or another primary source establishing such conduct. The existence of a quantitative investment affiliate and a sharp market reaction does not supply the missing evidence.

The revised package does not repeat the theory as a provocative possibility. Skepticism should tighten a claim, not license a more dramatic one.

The defensible correction is that the cost narrative was routinely described beyond the scope of the primary technical report.

Mixture of experts changed compute, not the meaning of cost

DeepSeek V3 uses a mixture-of-experts architecture. A router selects a subset of expert networks for each token, so the system does not activate all 671 billion parameters every time.

That can reduce arithmetic relative to a dense model with the same total parameter count. It does not make the other parameters disappear.

Weights still need memory across the system. Tokens may need to move between devices or nodes. Routing can create hot experts. Training and serving still depend on precision, batch, sequence length, memory, network, kernels, compiler, runtime, utilization, and operations.

[[Mixture of Experts Explained]] owns that technical distinction.

One attractive response was not a model evaluation

Dalton compared a DeepSeek output with a Gemini output and preferred aspects of DeepSeek's readable Markdown structure. That was a valid interface observation for one prompt.

It was not evidence of general accuracy, safety, grounding, privacy, latency, cost, multilingual performance, long-context behavior, tool use, or production suitability.

A real comparison would preserve exact model identities, dates, prompts, settings, tools, test cases, scoring, repetitions, failures, tokens, latency, cost, and reviewers. It would test the deployment path the organization plans to use.

The NIST AI Risk Management Framework adds govern, map, measure, and manage functions that extend far beyond benchmark or prompt impressions.

Efficiency does not settle hardware demand

E055 offered a plausible scenario: if each unit of compute becomes more useful, people may consume more of it.

That can happen. It is not guaranteed.

Hardware demand also extends beyond accelerator count. Memory, networking, storage, power, cooling, facilities, software, utilization, supply, and switching costs determine the delivered system.

NVIDIA's current SEC filings are the primary company source for its reported revenue, customers, competition, supply, and risk. They do not support an automatic investment conclusion.

Export-control claims have already moved

The episode discussed H800 hardware and U.S. restrictions in early 2025.

Current operative treatment requires the current Export Administration Regulations. BIS Part 742, checked July 28, 2026, contains current controls and license-review provisions. BIS also announced a January 2026 change for certain semiconductor exports to China under specified conditions.

That change demonstrates why a podcast-era explanation cannot decide a current transaction. Classification, destination, end user, ultimate parent, performance, use, license exception, review policy, and other facts require current counsel.

The surviving lesson is disciplined scope

E055's best instinct survives: read the technical claim before turning it into a company, market, or investment story.

Name the artifact. Name the training stage. Keep compute separate from dollars. List exclusions. Treat author benchmarks as author claims. Test exact models under exact conditions. Map total systems rather than counting GPUs. Use current filings and current law for current decisions.

That approach is less dramatic than a six-million-dollar revolution or a manipulation theory. It is also far more likely to remain useful.

The original episode is available on Spotify and YouTube. It preserves the February 2025 reaction and should not be used as current model, financial, market, legal, or export-control advice.

This page was developed with AI assistance from the raw E055 transcript and linked primary sources, then structured for human factual, technical, financial, legal, and editorial review. It is not an investment recommendation or export-control determination.