DeepSeek R1 Explained — Open Reasoning Model

Photo: Tara Winstead / Pexels
DeepSeek-R1 is an open-weights reasoning large language model released on January 20, 2025 by DeepSeek, a Hangzhou-based AI lab backed by the quantitative trading firm High-Flyer. It was the first openly released model to match OpenAI's o1 on hard math, coding, and logic benchmarks, and it shipped under a permissive MIT license, with weights, a technical paper, and a family of distilled smaller models all made public at once. What happened in the week after release changed how the entire industry talks about the cost of frontier AI.
On January 27, 2025, one week after launch, DeepSeek's mobile app hit the top of Apple's U.S. App Store free chart, ahead of ChatGPT. That same day, Nvidia's stock fell roughly 17% in a single session, wiping out about $589 billion in market capitalization — at the time, the largest single-day market-cap loss by any company in U.S. stock market history, according to Reuters. The trigger was a claim buried in DeepSeek's own paper: the final training run for R1's base model reportedly cost around $5.6 million in compute. If frontier-level reasoning no longer required nine-figure training budgets, the market's assumptions about who would win the AI race, and how many GPUs they would need to do it, suddenly looked wrong.
This article covers what R1 actually is, how DeepSeek trained it, why the market reacted the way it did, and what has happened to the DeepSeek model family since, including the R1-0528 update, DeepSeek-V3.2's sparse-attention breakthrough, and the V4 preview released in April 2026.
What DeepSeek-R1 actually is
DeepSeek did not release one model. It released a lineup, and understanding the difference between the pieces matters more than most coverage at the time made clear.
- DeepSeek-R1 — the flagship. A 671-billion-parameter Mixture-of-Experts (MoE) model, of which only about 37 billion parameters activate per token. Trained with a mix of reinforcement learning and a small amount of supervised fine-tuning to keep outputs readable.
- DeepSeek-R1-Zero — the more interesting research artifact. Same base architecture, but trained with pure reinforcement learning and no supervised fine-tuning at all. Reasoning behavior emerged on its own, without a single human-labeled "here's how to think" example, though R1-Zero's outputs sometimes mixed languages mid-sentence or skipped logical steps, which is why DeepSeek layered a lighter SFT pass on top to produce the more polished R1.
- Distilled models — six smaller open models (Qwen 1.5B, 7B, 14B, and 32B, plus Llama-3.1 8B and Llama-3.3 70B) fine-tuned on reasoning traces generated by R1 itself. The Qwen-32B distill matched OpenAI's o1-mini on several math and coding benchmarks despite being a fraction of the size, according to DeepSeek's technical report published on arXiv.
The distilled models made a point that mattered more than the flagship's benchmark scores: reasoning ability can be transferred. A small model taught with high-quality reasoning traces from a larger reasoning model can punch well above its parameter count. That single finding reshaped how open-source labs approached "how do we make small models smarter" for the rest of 2025.
The mechanism: reinforcement learning with verifiable rewards
Most aligned language models before R1 were shaped with RLHF (reinforcement learning from human feedback), where human raters compare pairs of model outputs and a reward model learns to approximate that human preference. It works, but it is expensive to collect, noisy to train on, and the reward model itself can be gamed by outputs that look good rather than are good.
Think of RLHF like training a chef by having food critics rate finished dishes. The critics' taste is subjective, inconsistent from one to the next, and expensive to collect at scale. DeepSeek trained R1 differently: for math, the "critic" is not a person, it is the answer key. For code, the "critic" is not a person, it is whether the code actually passes the unit tests. This is reinforcement learning from verifiable rewards (RLVR). The reward signal is binary and objective: right or wrong, pass or fail. No human preference judgment needed in the core training loop.
To make this stable at scale, DeepSeek used a policy-optimization method called GRPO (Group Relative Policy Optimization). Standard reinforcement learning approaches like PPO (Proximal Policy Optimization) require training a separate "critic" network that estimates how good a given state is, which roughly doubles the memory and compute cost of the setup. GRPO skips that. Instead, for a given prompt, the model samples a group of candidate answers, scores them against the verifiable reward, and treats the relative ranking within that group as the training signal. It is the difference between hiring a full-time judge to score every attempt in isolation versus just running several attempts side by side and rewarding whichever ones did comparatively better.
What is chain-of-thought reasoning, and why did it emerge on its own?
Chain-of-thought is the practice of a model writing out intermediate reasoning steps before producing a final answer, rather than jumping straight to a conclusion. In R1-Zero, DeepSeek's researchers observed the model's chain-of-thought length and reasoning depth increase organically over the course of RL training, purely because longer, more careful reasoning chains got rewarded more often with correct final answers. Nobody hand-wrote a "think step by step" instruction into the training data; the behavior was discovered because it worked. Our companion piece on AI reasoning and chain of thought goes deeper into how this generalizes across model families, including OpenAI's o-series and Gemini's thinking mode.
What is test-time compute, and how does R1 use it?
Test-time compute (also called inference-time compute) means letting a model spend more computation at the moment you ask it a question, rather than only during training. R1 can generate long internal reasoning traces, sometimes thousands of tokens, before committing to an answer, effectively trading response latency for accuracy on hard problems. The R1-0528 update, described below, pushed this further: DeepSeek reported the updated model uses close to double the "thinking" tokens on hard problems compared to the original R1 (about 23,000 tokens versus roughly 12,000), which is a direct driver of its higher benchmark scores.
Benchmark performance at launch
R1 matched or beat OpenAI's o1 on most reasoning benchmarks when it launched in January 2025, based on figures published in DeepSeek's arXiv technical report.
Table: DeepSeek-R1 vs. contemporary models at January 2025 launch
| Benchmark | DeepSeek-R1 | OpenAI o1 | Claude 3.5 Sonnet | Llama 3.3 70B |
|---|---|---|---|---|
| AIME 2024 (competition math) | 79.8% | 79.2% | 16.0% | 23.3% |
| MATH-500 | 97.3% | 96.4% | 78.3% | 73.8% |
| GPQA Diamond (graduate-level science) | 71.5% | 75.7% | 65.0% | 50.5% |
| MMLU (general knowledge) | 90.8% | 91.8% | 88.3% | 86.0% |
| Codeforces (competitive programming, Elo) | 2029 | 2061 | 717 | 759 |
| Open weights? | Yes (MIT) | No | No | Yes (custom license) |
Source: DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning," arXiv:2501.12948, January 2025.
The benchmark story is close. The cost story is not. R1's API pricing at launch was $0.55 per million input tokens and $2.19 per million output tokens, versus OpenAI o1's pricing at the time of roughly $15 per million input tokens and $60 per million output tokens, putting R1 at roughly a 27x discount for broadly comparable reasoning quality on math and code. That combination, near-parity performance at a fraction of the price, is what made R1 spread through developer communities within days of release rather than the usual slow adoption curve for a new open model.
Why the market reacted so hard
Three separate threads converged in the same week and amplified each other.
First, the $5.6 million training-cost claim suggested that frontier-level capability did not require the $100 million-plus training runs that OpenAI, Anthropic, and Google were reportedly spending. If that were generally true, the enormous capital expenditure U.S. labs and cloud providers had committed to, largely in the form of Nvidia GPU orders, looked less clearly justified.
Second, R1 shipped as open weights under an MIT license. Any company, anywhere, could download it and run it on their own hardware, with no API dependency on DeepSeek at all. That undercut the assumption that frontier reasoning capability was a moat only a handful of well-funded labs could defend.
Third, a Chinese lab had beaten every U.S. lab to the punch on open reasoning models, and it had done so despite U.S. export controls specifically designed to restrict Chinese access to the most advanced Nvidia AI chips. That was the geopolitical sting: the controls were supposed to slow this exact outcome down.
It's worth being precise about what the $5.6 million figure actually measured, because it was widely misquoted in the days after release. DeepSeek's paper described that number as the cost of the final pre-training run only, calculated at roughly 2.78 million H800 GPU-hours at an assumed $2 per GPU-hour. It explicitly excluded prior research, failed experimental runs, data collection, salaries, and the underlying infrastructure investment. DeepSeek's total R&D spend was almost certainly a good deal higher than $5.6 million. But the more nuanced version of the story didn't stop the headline number from moving markets. Nvidia's stock partially recovered over the following weeks as some analysts argued that cheaper inference would expand total AI demand rather than shrink it, an argument based on the economic idea known as the Jevons paradox: when a resource becomes more efficient to use, total consumption of it often rises rather than falls.
How DeepSeek did it with restricted GPUs
DeepSeek trained R1's base model using Nvidia H800 GPUs, the export-compliant, deliberately bandwidth-limited variant of the H100 that Nvidia sells into the Chinese market to comply with U.S. export controls. DeepSeek's paper describes using a cluster of roughly 2,048 H800s for the main training run. Several architectural and engineering choices let the team squeeze more useful compute out of hardware that was intentionally hobbled compared to what U.S. labs had access to.
- Mixture-of-Experts architecture. Only 37 billion of R1's 671 billion total parameters activate for any given token, which dramatically cuts the floating-point operations needed per forward pass compared to a dense model of similar total size. Our Mixture of Experts explained article covers how routing decides which "experts" handle which tokens, and why this lets trillion-parameter-class models run at a fraction of the inference cost of a dense model that size.
- FP8 mixed-precision training. Using 8-bit floating point numbers for most training computation roughly halves memory use compared to the FP16 precision most labs used at the time, at the cost of needing careful numerical engineering to avoid instability.
- Custom low-level communication kernels, written to work around the H800's deliberately reduced inter-GPU interconnect bandwidth relative to the export-restricted H100.
- Multi-head latent attention (MLA), a technique that compresses the key-value cache (the memory a model uses to "remember" earlier tokens in a conversation while generating new ones), cutting the memory footprint needed to serve long contexts.
DeepSeek's technical paper was also unusually detailed by the standards of frontier AI publishing in 2025. Most U.S. labs by that point had scaled back how much architectural and training detail they disclosed publicly, citing competitive and safety concerns. DeepSeek published enough for other labs and independent researchers to attempt real reproductions of parts of the training pipeline within weeks, which is part of why R1's technical claims were scrutinized and largely corroborated rather than dismissed as marketing.
Geopolitical implications: the export-control debate
R1 reframed a specific policy bet the United States had been making since at least 2022: that restricting Chinese labs' access to the most advanced AI accelerators would meaningfully slow Chinese frontier AI progress. R1 demonstrated that algorithmic and systems-engineering efficiency could partly substitute for raw compute access, at least for this generation of models. That is a narrower claim than "export controls don't work," but it was enough to trigger a fast policy response.
Within weeks of R1's release: U.S. policymakers proposed tightening chip export rules further and discussed restricting DeepSeek's app on government devices; multiple countries, including Italy, Taiwan, and Australia, moved to restrict or ban DeepSeek's hosted service on government or state-owned systems, citing data-privacy concerns tied to Chinese data-security law rather than the model's capabilities; and U.S. labs visibly accelerated their own reasoning-model roadmaps, with OpenAI's o3, Anthropic's Claude 3.7 Sonnet with extended thinking, and Google's Gemini 2.5 Pro all arriving in the months that followed. Whether that acceleration would have happened on the same timeline without R1's release is not something anyone can prove, but the sequencing was hard to miss.
The DeepSeek model family since January 2025
R1's January 2025 release was the beginning of a rapid cadence of updates, not the end of the story. Anyone evaluating DeepSeek's models today needs to know what has shipped since, because "DeepSeek-R1" as most people used it by mid-2026 was not the exact model that made headlines.
Table: DeepSeek model release timeline, January 2025 to April 2026
| Release | Date | What changed |
|---|---|---|
| DeepSeek-R1 / R1-Zero | January 20, 2025 | Original launch. MIT license, 671B MoE, RLVR + GRPO training. |
| DeepSeek-R1-0528 | May 28, 2025 | Mid-generation update to the same R1 architecture. AIME 2025 score rose from 70.0% to 87.5%; GPQA Diamond rose from 71.5% to 81.0%; roughly double the average "thinking" tokens per hard query. |
| DeepSeek-V3.1 / V3.1-Terminus | Mid-2025 | Interim base-model update ahead of the V3.2 architecture change. |
| DeepSeek-V3.2-Exp | September 29, 2025 | Introduced DeepSeek Sparse Attention (DSA), a fine-grained sparse-attention mechanism that cuts long-context inference cost. API pricing dropped by more than half. |
| DeepSeek-V3.2 / V3.2-Speciale | February 15, 2026 | Full non-experimental V3.2 release. The high-compute Speciale variant reportedly reached gold-medal-level performance on international math and programming olympiad-style benchmarks. |
| DeepSeek-V4-Pro / V4-Flash (preview) | April 24, 2026 | Preview of DeepSeek's next-generation architecture: 1.6T-parameter Pro and 284B-parameter Flash variants, both with combined "thinking" and "non-thinking" modes and a 1-million-token context window. |
Sources: DeepSeek API documentation news archive; OpenRouter model benchmark pages for R1-0528, V3.2, and V3.2-Exp; Hugging Face model cards for deepseek-ai/DeepSeek-R1-0528.
What changed in the R1-0528 update?
DeepSeek released R1-0528 on May 28, 2025, as a mid-generation refresh of the same underlying R1 architecture rather than a new model family. The update meaningfully closed R1's remaining benchmark gaps: AIME 2025 math scores jumped from 70.0% to 87.5%, LiveCodeBench coding scores rose by nearly 10 points, and DeepSeek added support for structured JSON output and function calling, both of which the original R1 lacked and which developers had been asking for since launch. The update also reduced hallucination rates according to DeepSeek's release notes, though independent, large-scale hallucination benchmarking of the update was more limited than the math and coding evaluations.
What is DeepSeek Sparse Attention, and why does it matter?
Standard "dense" attention in a transformer model compares every token in a sequence against every other token, which is why cost grows quadratically as context length increases; doubling the context roughly quadruples the attention computation. DeepSeek Sparse Attention (DSA), introduced in V3.2-Exp in September 2025, selects only a relevant fraction of prior tokens to attend to for each new token generated, rather than all of them. It is closer to how a reader skims back through a long document for the specific paragraphs that matter instead of rereading every page. DeepSeek reported this cut long-context inference costs enough to justify halving V3.2's API price compared to V3.1-Terminus, according to DeepSeek's official API documentation and coverage from VentureBeat.
Has DeepSeek-R2 been released?
As of this update, no. DeepSeek-R2 has been persistently rumored since early 2025, and multiple outlets reported conflicting claims about a possible launch in early-to-mid 2026, but no official DeepSeek-R2 model card, weight release, or API listing had appeared as of mid-2026. Reuters reported that DeepSeek's CEO, Liang Wenfeng, held back an R2 release because he was reportedly not satisfied with its performance relative to competing models. What actually shipped in that window instead was the V3.2 family in February 2026 and the V4-Pro/V4-Flash preview in April 2026, both under DeepSeek's now-standard MIT license.
How DeepSeek-R1 compares to today's reasoning models
By mid-2026, R1 is no longer competing against a single OpenAI model; it sits in a crowded field of reasoning systems that all launched in response to, or roughly alongside, it. Meta's open-weight Llama 4 models are another prominent open alternative.
Table: Reasoning model landscape, mid-2026 snapshot
| Model | Developer | Weights | Context window | Distinguishing trait |
|---|---|---|---|---|
| DeepSeek-R1-0528 | DeepSeek | Open (MIT) | 128K tokens | Cheapest frontier-class reasoning per token; fully self-hostable. |
| DeepSeek-V3.2-Speciale | DeepSeek | Open (MIT) | Long-context, DSA-optimized | Olympiad-level math/code performance at high compute settings. |
| OpenAI o3 | OpenAI | Closed | Varies by tier | Deep integration with OpenAI's broader product and tool-use ecosystem. |
| Claude 3.7 Sonnet (extended thinking) | Anthropic | Closed | 200K tokens | Adjustable "thinking budget" exposed directly to developers. |
| Gemini 2.5 Pro | Google DeepMind | Closed | 1M tokens | Native multimodal reasoning across text, image, audio, and video. |
Closed models generally still lead on some general-knowledge and multimodal benchmarks. What R1 and its successors changed permanently is the price floor: once a genuinely competitive reasoning model was available open-weight and dramatically cheaper per token, every closed-model provider had to defend its pricing with something other than "there is no open alternative," because by 2026 there clearly was.
How to actually use DeepSeek's models
- Web or app: chat.deepseek.com and the DeepSeek mobile app, both free to use. Conversations are processed on servers located in China and are subject to Chinese data-handling law, which several governments have flagged as a reason to restrict use on official devices.
- API: api.deepseek.com, using an OpenAI-compatible request format, which makes swapping DeepSeek into an existing OpenAI-based codebase a comparatively small change. Note that DeepSeek retired the old
deepseek-chatanddeepseek-reasonerAPI identifiers after July 24, 2026, routing both to the V4-Flash model's non-thinking and thinking modes respectively; anyone with older integrations should check current DeepSeek API documentation before assuming legacy endpoint names still work. - Self-hosting: model weights for the full R1 and V3.2 families are downloadable from Hugging Face. The full 671B/685B-class models need roughly the equivalent of eight H100-class GPUs running FP8 precision to serve at reasonable latency; the distilled 7B–32B models run comfortably on a single consumer GPU with 24 GB of VRAM or less, depending on quantization.
- Third-party hosting on non-Chinese infrastructure: providers including Together AI, Fireworks AI, Groq, and Perplexity host DeepSeek's open-weight models on U.S. or EU infrastructure, which sidesteps the data-residency concern entirely for teams that need the model's capability without routing traffic through China-based servers.
What people get wrong about DeepSeek-R1
The most common misconception is treating "$5.6 million to build a frontier model" as a literal, all-in cost comparison against what OpenAI or Anthropic spend. It is not. The figure is a single training run's GPU-hour cost, explicitly excluding R&D, failed experiments, data pipeline construction, and salaries, by DeepSeek's own description in its paper. A fair comparison would need the same accounting standard applied to closed-lab training costs, which none of them publish in comparable detail, so a clean apples-to-apples number does not currently exist.
The second common error is calling R1 "fully open source" without qualification. DeepSeek released the model weights under the MIT license, which is genuinely permissive: you can run, modify, and commercially deploy the weights with minimal restriction. But DeepSeek did not release the training data or the complete training code, which means the model doesn't meet the stricter definitions of open source used by organizations like the Open Source Initiative. "Open weights" is the more accurate term, and it is the term used through the rest of this article for exactly that reason.
A third misconception, common in early 2025 coverage, treated R1 as a single static model rather than the starting point of an active release cadence. By mid-2026, DeepSeek had shipped R1-0528, two V3.1 variants, V3.2-Exp, the full V3.2 and V3.2-Speciale release, and a V4 preview, each with materially different benchmark scores and pricing. Any comparison of "DeepSeek versus [other model]" that doesn't specify which DeepSeek release it means is comparing against a moving target.
Finally, some coverage conflated DeepSeek's chip usage with a demonstration that export controls categorically failed. What R1 actually showed is narrower: a well-resourced, highly skilled engineering team can extract more capability per GPU-hour than assumed, using architecture and training-algorithm choices rather than raw hardware scale. That is a real and consequential finding about algorithmic efficiency. It is not proof that compute access stops mattering, and DeepSeek's own roadmap since R1, larger models, larger contexts, more total training compute, suggests the lab itself does not believe that either.
Frequently Asked Questions
Q: When was DeepSeek-R1 released? A: January 20, 2025, by DeepSeek, a Hangzhou-based AI lab backed by the quantitative trading firm High-Flyer.
Q: Is DeepSeek-R1 really open source? A: The model weights are released under the MIT license, which is permissive and allows commercial use. The training data and full training code are not released, so it is more accurately described as "open weights" rather than fully open source in the strict Open Source Initiative sense.
Q: Did training DeepSeek-R1 really cost only $5.6 million? A: That figure, from DeepSeek's own technical paper, refers only to the final pre-training compute cost, roughly 2.78 million H800 GPU-hours at an assumed $2 per hour. It excludes prior research, failed training runs, salaries, and infrastructure. The true total R&D cost was almost certainly higher, but DeepSeek has not published that broader figure.
Q: Why did Nvidia's stock drop because of DeepSeek? A: Investors feared that if frontier AI models required less compute than assumed, demand for Nvidia's high-end GPUs would grow more slowly than priced in. Nvidia fell about 17% on January 27, 2025, a roughly $589 billion single-day market-cap loss, per Reuters. The stock partially recovered in subsequent weeks as some analysts argued cheaper inference would expand total AI demand rather than shrink it.
Q: How does DeepSeek-R1 compare to OpenAI's o1? A: At launch, R1 matched or slightly trailed o1 on math and coding benchmarks and trailed slightly on general knowledge (MMLU). R1 was roughly 27 times cheaper per token and shipped as open weights; o1 was closed and available only through OpenAI's API. The May 2025 R1-0528 update closed most of the remaining gap on math and reasoning benchmarks.
Q: What is DeepSeek-R1-0528, and is it different from the original R1? A: R1-0528 is a May 28, 2025 update to the same R1 architecture, not a new model. It substantially improved math benchmark scores (AIME 2025 rose from 70.0% to 87.5%), coding scores, and added JSON output and function-calling support that the original R1 lacked.
Q: Has DeepSeek released R2? A: Not as of mid-2026. Reuters reported the release was held back over performance concerns. DeepSeek instead shipped the V3.2 and V3.2-Speciale models in February 2026 and previewed a next-generation V4-Pro/V4-Flash architecture in April 2026.
Q: Is it safe to use DeepSeek's hosted chat service? A: The hosted chat.deepseek.com service sends conversations to servers in China, subject to Chinese data-security law. Several governments, including Italy, Taiwan, and Australia, have restricted its use on official devices for that reason. For sensitive data, use a version hosted on non-Chinese infrastructure (Together AI, Fireworks, Perplexity) or self-host the open weights.
Q: Can I run DeepSeek-R1 on my own computer? A: The full 671B-parameter R1 model needs roughly eight H100-class GPUs to serve efficiently, well beyond consumer hardware. The distilled versions (Qwen 1.5B through 32B, Llama 8B and 70B) are far smaller and can run on a single consumer GPU, with the smallest variants runnable on a laptop-class GPU using quantization.
Q: What is Mixture-of-Experts, and why does DeepSeek use it? A: Mixture-of-Experts is an architecture where a model has many specialized sub-networks ("experts") but only activates a small subset for any given input, rather than running the full parameter count every time. R1 has 671 billion total parameters but activates only about 37 billion per token, which cuts inference cost dramatically compared to a dense model of the same total size. See our Mixture of Experts explained article for the full mechanism.
The takeaway
DeepSeek-R1 mattered less because of any single benchmark score and more because it broke an assumption: that frontier-grade reasoning capability required the compute budgets only a handful of American labs could afford. The reality that emerged over the following eighteen months was more nuanced than the January 2025 headlines suggested, DeepSeek kept spending, kept training larger models, and kept iterating (R1-0528, V3.1, V3.2, V4-Preview) at a pace that looks a lot like every other frontier lab's roadmap. What changed permanently is the price floor for reasoning-capable AI and the fact that anyone can now download comparable capability and run it themselves. That is the part of the DeepSeek story that outlasted the stock-market headline.
Read next:
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles

Gemini 2.5 Pro Explained — Google's Best Model
Gemini 2.5 Pro is Google DeepMind's reasoning model with a 1M-token context and top scores on coding and reasoning benchmarks at launch.

Tool Use vs Agent — What's the Difference?
Tool use lets an LLM call one function. An agent runs a loop of reasoning and tool calls until done. Here's the real distinction.

Llama 4 Explained — Meta's Open AI Models
Llama 4 is Meta's April 2025 open AI family — Scout, Maverick, and the shelved Behemoth — with native multimodal input and a 10M-token context window.

Open Source AI vs Proprietary AI: The Real Tradeoffs
Open source AI vs proprietary AI compared: what licenses allow, how far the capability gap has narrowed, and when self-hosting actually beats a paid API.

What Is GPT-5? OpenAI's Latest Model
GPT-5 explained: OpenAI's unified model with auto-routing, three sizes, benchmarks vs Claude 4, Gemini 2.5 and Grok 4, pricing, and what changed.

What Is Mistral AI? Europe's LLM Lab, Explained
Mistral AI is the Paris lab racing to give Europe its own frontier LLM. How its models, open-weight strategy, and ASML-backed funding compare to US rivals.