What the 'Notes on DeepSeek' Essay Gets Right About Open-Weights Economics

On June 10, 2026, a Twitter thread titled "Notes on DeepSeek" hit the Hacker News front page with 95 points and 72 comments. The author, Niko McCarty, had visited DeepSeek headquarters in Hangzhou the previous Tuesday. The post was brief - field notes more than analysis - and the original tweet was deleted before the HN thread finished warming up. A commenter pasted the full text anyway, because the internet works that way, and the discussion that followed was one of the more substantive conversations about AI economics that HN has produced in months.

Last updated: June 10, 2026

What the Essay Actually Says

McCarty's observations are deceptively simple. DeepSeek operates out of an unmarked 12-story building in Hangzhou. There is no company branding visible from the street. When he asked why, the team said: "Well, there are many companies in this building, and we are not special." The company has roughly 300 employees - by McCarty's count, at least an order of magnitude smaller than Anthropic. The Head of Infrastructure appeared to be around 30 years old.

The team reads western AI writers. They listen to Dwarkesh. They read Gwern. They had never met anyone from Anthropic. They were not worried about hostile AGI scenarios. Their main concern was job loss among young people in China, where youth unemployment is already elevated.

The line that generated the most discussion on HN: "As a whole, China seems to treat AI as just another technology, rather than as some kind of singularity moment. National attention is still on basic needs and infrastructure buildouts, and on providing more medicines for people. The 'dreams of singularity seem like a luxury or distant consideration.'"

That observation has real implications - not primarily as geopolitical commentary, but as a signal about how competitive incentives are structured in Chinese AI development versus American AI development.

The Pareto Frontier Point

McCarty's essay is thin on model economics, but the HN thread filled in the gap quickly. One of the most-upvoted observations: "DeepSeek models are on the Pareto frontier of cost/performance. That's the far more important one than just making a top-scoring model."

This is the crux of why open-weights economics matter for working developers, and it connects directly to what we have covered in our DeepSeek V4 posts here on Developers Digest. DeepSeek V4 Flash is not the best model on every benchmark. It does not need to be. What it does is sit at a price-performance point that disciplines every other API on the market - including the ones that charge 10x more per token for marginal quality gains on tasks that do not require frontier capability.

One developer in the HN thread put it plainly: "It's crazy how much you get out from DeepSeek V4 Flash alone. Ever since I found Opencode Go, AI coding is fun again. I always hated the feeling of working inside a fenced constraint where if I just go hard enough I suddenly hit a wall and have to pay up a lot more."

Another: "If it wasn't for China, I would probably have to spend $100/month on AI instead of $10 like I do currently while using DeepSeek and MiMo."

These are not abstractions. They describe a concrete change in what developers can afford to run continuously, at scale, without worrying about hitting a ceiling.

Where the HN Discussion Pushed Back

The essay is not without problems, and the HN thread identified them fairly.

Several commenters noted that McCarty's read of DeepSeek as humble and straightforwardly focused on infrastructure elides some complexity. One wrote that the piece "felt void of actual information, with the restaurant replaced by the office" - a fair critique of first-impression journalism. Another called it "a puff piece written by someone who didn't know (or didn't care) they were being managed." When a company tells a visiting journalist "we are not special" while building one of the most strategically significant AI labs on the planet, a healthy dose of skepticism is warranted.

The more substantive pushback concerned distillation and originality. Some commenters argued that DeepSeek's efficiency gains rest partly on distillation from frontier US models, which raises questions about how far that can go independently. Others found this unpersuasive - pointing out that distillation is a standard technique, that the innovations in DeepSeek's architecture (mixture-of-experts training efficiency, in particular) are documented and real, and that the irony of Anthropic or OpenAI complaining about training data provenance is significant.

The open-weights-as-security-risk angle also surfaced. One technically-grounded commenter pushed back on the claim that open weights enable adversarial state-level capabilities: "It's trivial for me to download one of their models and run it on my Spark, and there are all sorts of ways to strip out their Tiananmen-denialism or whatever." The argument cuts both ways - openness is a feature and a risk simultaneously, and where you land depends heavily on your threat model.

What is less contested: whether DeepSeek and its peers are innovating, or whether they matter for the price of compute accessible to independent developers. On both counts, the evidence is clear.

The Ecosystem the Essay Describes

McCarty notes that competition is coming from Alibaba's Qwen, ByteDance, and Moonshot AI's Kimi. This is accurate as of April 2026, when DeepSeek V4, Kimi K2.6, and Qwen 3.6 released within hours of each other - an event that Graham Webster at Stanford's DigiChina Project described as demonstrating that "rumors of the death of Chinese open-weight AI advancement have been greatly exaggerated."

That cluster of simultaneous releases is the structural story underneath McCarty's field notes. It is not one company with a cost advantage. It is a cohort of labs - DeepSeek, Qwen, Kimi, GLM, MiniMax - releasing competitive open-weight models on permissive licenses, on a cadence that tracks the US frontier at a significant discount. This is the cost floor that matters for developers. No single release locks in pricing; the floor is maintained by competition within the cohort itself.

The practical consequence for routing decisions: the question is no longer "open source vs. closed" as a binary. It is a layered question. Which tasks require frontier judgment? Which tasks are volume operations where a Pareto-efficient open model covers 90% of quality at 10% of cost? The DeepSeek V4 Flash / Pro split that we covered in our developer guide is a working example of this - Flash for throughput-heavy agentic loops, Pro for the cases where inference quality is the bottleneck.

The Singularity Contrast as Competitive Advantage

The observation that China treats AI as "just another technology" is one McCarty offers somewhat neutrally. The HN thread read it as a meaningful signal, and so do we.

US frontier labs have built their brand positioning - and to some degree their internal culture - around the claim that they are building something categorically different from prior technology. That framing has genuine strategic value: it attracts talent, justifies valuation, and creates regulatory moats. It may also be true, depending on which trajectory you believe in.

But it creates a pricing and product distortion. When your model is a step toward AGI, you price accordingly and you guard access accordingly. When your model is infrastructure - like a database, like a CDN - you compete on reliability, latency, and cost per token.

One HN commenter, quoting the essay's singularity line, wrote: "From the notes, this part sat with me as the real difference. I would argue the US providers have gone full tilt into sales culture with respect to AI. The constant release cycles of things that don't exist for most people, the gatekeeping - it's all a part of large brand toxic sales and marketing."

That may be overstated. But the underlying observation - that the cultural framing of AI in the US shapes pricing behavior - is not obviously wrong. A lab that believes it is building God has different incentives than a lab that believes it is building a faster database. The second lab competes harder on price.

The Practical Developer Takeaway

The essay and the discussion around it converge on something actionable. Open-weight frontier-adjacent models from DeepSeek, Qwen, Kimi, and MiniMax are not a geopolitical story for most developers - they are a cost floor. That floor disciplines API pricing across the entire market, including from labs whose models you prefer for quality reasons.

The routing pattern that follows from this: use open models for the tasks where throughput matters and quality tolerances are wide - code generation scaffolding, document summarization, classification, batch enrichment. Reserve frontier closed models for the tasks where marginal quality improvement has asymmetric value - user-facing inference, complex multi-step reasoning, judgment calls with downstream consequences.

This is not a novel idea, but the essay's ground-level view of DeepSeek reinforces why it is durable. A 300-person company in an unmarked building, staffed by people in their late twenties, is not going away. Neither are the labs competing with it in the same ecosystem. The cost floor they collectively enforce is structural, not a temporary price war.

Whether that is entirely good news is a fair question. The HN thread explored the security dimensions, the distillation ethics, and the geopolitics at length, without resolution. But for developers deciding what to pay per million tokens this month, and next month, the direction is clear.

---

FAQ

What is the "Notes on DeepSeek" essay?

"Notes on DeepSeek" is a short first-person account posted to Twitter/X by writer Niko McCarty, describing a visit to DeepSeek's headquarters in Hangzhou, China in June 2026. It was shared on Hacker News on June 10, 2026, where it reached 95 points and 72 comments. The original post was deleted but preserved in the HN thread. Key observations include DeepSeek's small headcount (around 300 employees), its low-profile operations, and its team's pragmatic view of AI as technology rather than a civilization-altering singularity.

How do open-weight models from DeepSeek affect API pricing?

Open-weight models create a cost floor across the AI API market. When a capable model is available for self-hosting or through affordable third-party inference providers, closed-API providers face pricing pressure even on their own products. DeepSeek V4 Flash, for example, delivers competitive performance on a wide range of developer tasks at a price point significantly below frontier closed models, which compresses the justifiable premium for those models on non-judgment tasks.

Is DeepSeek's efficiency based on original research or distillation?

Both, by most accounts. DeepSeek has published research on mixture-of-experts architectures and training efficiency techniques that are independently documented. The distillation critique - that some of DeepSeek's capability gains come from training on outputs of US frontier models - is raised periodically, but critics note that this technique is industry-wide and does not account for the architectural innovations DeepSeek has demonstrated. The debate in the HN thread did not reach consensus, and remains an open question worth following.

Should I route all my AI workloads through open-weight models?

Probably not all of them. The routing pattern most developers are converging on is task-based: open-weight or frontier-adjacent models for high-volume, lower-stakes tasks (batch processing, summarization, code scaffolding); closed frontier models for tasks where marginal quality gains have significant downstream value (complex reasoning, user-facing generation, judgment-heavy classification). DeepSeek V4's Flash/Pro split is a practical example of this within a single model family. The goal is matching model cost to task requirements, not minimizing spend across the board.

---

Sources

• Niko McCarty, "Notes on DeepSeek" (Twitter/X, June 10, 2026) - original post deleted; text preserved in HN comments: https://news.ycombinator.com/item?id=48476474
• Hacker News thread, "Notes on DeepSeek" (95 points, 72 comments, June 10, 2026): https://news.ycombinator.com/item?id=48476474
• Graham Webster, "Death of Chinese Open-Weight AI Exaggerated," Here It Comes / Transpacifica (April 24, 2026): https://herecomes.transpacifica.net/p/death-of-chinese-open-weight-ai-exaggerated
• Stanford HAI / DigiChina, "Beyond DeepSeek: China's Diverse Open-Weight AI Ecosystem and Its Policy Implications": https://hai.stanford.edu/assets/files/hai-digichina-issue-brief-beyond-deepseek-chinas-diverse-open-weight-ai-ecosystem-policy-implications.pdf