Future Horizons

The Open-Source Reckoning: How Freely Available AI Models Are Redrawing the Map of Power

Open weights became the cheaper half of the market: DeepSeek's V4, Mistral's 3 billion euro Series D, the January 2026 H200 rule, the EU AI Act's open-source carve-out, and the dual-use risk none of it solves.

Gabriele Masetti · 2026-09-16

The Moat That Disappeared

For the first decade of the modern AI era, one assumption about the structure of power in artificial intelligence went largely unchallenged: the most capable systems required so much data, so much compute and so much specialised expertise that only a few well-capitalised organisations could build them. OpenAI, Google DeepMind, Anthropic and a handful of others held what looked like an insurmountable structural advantage. No open-source community, no university group, no government outside the wealthiest could compete. The moat appeared permanent.

In January 2025, a Chinese AI research lab called DeepSeek published a paper describing a reasoning model called R1. The paper reported that R1 achieved performance comparable to OpenAI's o1 model on a range of reasoning benchmarks — and that it had been trained for a reported $5–6 million, a fraction of the hundreds of millions that comparable Western models had cost. More importantly: the weights were released under an open-source licence.

The response in the AI industry was something between shock and recalibration. Within days, DeepSeek-R1 had been downloaded millions of times, integrated into dozens of applications, and fine-tuned by researchers around the world. Nvidia's stock dropped 17% in a single trading session — not because R1 was necessarily better than o1 in all respects, but because it demonstrated that the assumption of prohibitive training costs was wrong. The moat was not as deep as it had appeared.

DeepSeek-R1 was not a singular event. It was the most dramatic expression of a trend that had been building since Meta released Llama in 2023 and Llama 2 later that year: the systematic erosion of the performance gap between open and closed AI models, and the redistribution of AI capability from a handful of proprietary labs to the global research community.

The reckoning has arrived, and its implications for business models, geopolitics, national security and AI governance are only beginning to be understood.

The Economics of Open Versus Closed

Closed models — accessed through APIs provided by OpenAI, Anthropic, Google, and others — are priced per token: the unit of text the model processes. OpenAI retired GPT-4o from ChatGPT on 13 February 2026; its flagship as of September 2026, GPT-6 Astra, lists at $10 per million input tokens and $50 per million output on standard processing. DeepSeek serves V4-Pro — an MIT-licensed model anyone can download and run — through its own API at $0.66 to $1.32 per million input tokens on a cache miss and $1.98 to $3.96 per million output, depending on the hour. At peak rates that is roughly eight times cheaper on input and thirteen times cheaper on output, before anyone considers hosting the weights themselves. For organisations processing very large volumes of text — customer service automation, document analysis, content generation — this difference is transformative.

Beyond cost, open-source models offer several advantages that closed APIs cannot match:

Customisation: a closed API provides a fixed model that can be prompted but not fundamentally altered. An open-weight model can be fine-tuned on proprietary data, adjusted for specific domains, and modified at the architecture level. For highly specialised applications — medical diagnosis, legal document analysis, scientific research assistance — this customisability is often more valuable than raw benchmark performance.

Data privacy: using a closed API requires sending data to a third-party server. For organisations handling sensitive data — medical records, legal documents, financial information, classified government information — this is either legally prohibited or commercially unacceptable. Open-weight models can be run on-premises or in private cloud environments, keeping data within the organisation's control.

Independence: dependence on a closed API creates vendor lock-in and exposure to the API provider's pricing decisions, terms of service changes, and potential business failures. Open-weight models eliminate this dependency.

Auditability: a closed model is a black box — its user cannot inspect the training data, verify safety properties or audit decisions. An open-weight model is at least more auditable, and with open code and training data — which most open-weight releases still do not provide — fully auditable.

A 2024 analysis published in the Berkeley California Management Review estimated that the "optimal reallocation" from closed to open-source models, if fully adopted across the economy, could free up approximately $25 billion annually — capital that could be redirected from AI inference costs to complementary AI investments like training, fine-tuning, and application development.

The Performance Convergence

The economics of open-source AI would be irrelevant if open models were substantially less capable than closed ones. For most of 2022 and 2023, the performance gap was significant enough to matter for many applications. Since late 2023, the gap has been closing at a pace that has surprised even optimistic observers.

Meta's Llama 2 (July 2023) was competitive with GPT-3.5 but well behind GPT-4; Llama 3 (April 2024) closed much of that gap; Llama 3.1 (July 2024) was competitive with GPT-4o. Llama 4 (released 5 April 2025) matched or exceeded closed models on certain benchmarks and narrowed the gap substantially on others. It is also the last model to carry the Llama name. The model powering Meta AI since April 2026 is Muse Spark, from Meta Superintelligence Labs, which is not part of the Llama line; Meta shipped Muse Spark 1.1 on 9 July 2026. Then, on 10 August 2026, it published Muse Glimmer, a 30-billion-parameter agentic model distilled from Muse Spark and released under Apache 2.0 with a 131,072-token context — its first open-weight release since Llama 4, and small enough to run on a single consumer GPU. Mark Zuckerberg said the weights for Muse Spark 1.2, Meta's latest foundation model, would follow.

Mistral's trajectory was faster still: Mistral 7B (September 2023) showed that smaller, more efficiently trained models could approach much larger closed models on many tasks, and Mixtral 8x7B (December 2023) matched GPT-3.5 on most benchmarks while costing far less to run.

DeepSeek's contributions have been particularly significant because they have come with detailed technical disclosures. DeepSeek-V3 and DeepSeek-R1 not only provided competitive open-weight models but also described training methodologies — particularly around reinforcement learning from verifiable outcomes and mixture-of-experts architectures — that have been widely adopted by the research community.

Both have since been superseded by the lab's own work: on 24 April 2026 DeepSeek released V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion / 13 billion) on Hugging Face under an MIT licence, with a 1-million-token context as the default across its services and 80.6% on SWE-bench Verified for V4-Pro.

The centre of gravity in open weights also moved east. The two largest open-weight models of 2026 are Chinese: DeepSeek's V4-Pro and Moonshot AI's Kimi K3, at 2.8 trillion total parameters with 104 billion active and the same million-token window. Mistral remains the European anchor, and Meta's return to open weights came at 30 billion parameters rather than at the frontier, which leaves the 2024 framing — Western open source against Chinese proprietary AI — describing the wrong map.

The pattern is consistent: open-source models lag frontier closed models by 6–18 months on most benchmarks, and the lag has been shrinking — in some domains it has closed. Analysis from a16z suggests that open models achieve roughly 90% of closed-model performance at release and typically close the gap within 6–12 months.

Closed models still account for approximately 96% of AI revenue, suggesting that the commercial transition is in early stages. But the direction of travel is unmistakable, and the companies building business models on the assumption that closed models will always be substantially more capable are operating on a narrowing foundation.

DeepSeek as Geopolitical Event

DeepSeek-R1's release was not merely a technical and economic event. It was a geopolitical one.

First, it challenged the effectiveness of US export controls on advanced semiconductors. A central element of US AI strategy under both the Trump and Biden administrations was restricting China's access to frontier AI chips — Nvidia's A100, H100 and H200 GPUs among them — on the theory that compute superiority translates to AI capability superiority.

DeepSeek-R1 was reportedly trained on Nvidia H800 chips (a lower-performance export-legal variant) and used efficiency innovations — including distillation from larger models, aggressive quantisation, and novel training curricula — to achieve frontier-class performance with less compute than Western models. That called into question whether chip export controls could long remain effective as a mechanism for maintaining Western AI advantage.

Second, it demonstrated the depth of Chinese AI research capability. US policymakers and industry leaders had generally assumed that Chinese AI research was derivative of Western work, capable of implementation and application but not fundamental innovation. DeepSeek-V3 and R1 demonstrated otherwise: the training methodology papers described genuine innovations in reinforcement learning and mixture-of-experts architecture that have influenced subsequent work at Western labs.

Third, it transformed the geopolitics of open-source AI. Prior to DeepSeek, the most capable open models were primarily from Western organisations (Meta, Mistral, Stability AI), allowing the US and Europe to frame open-source AI as an extension of their AI leadership. Post-DeepSeek, the most influential open models include significant Chinese contributions, complicating any simple narrative of Western open-source versus Chinese proprietary AI.

The response in Washington was a mixture of alarm and reassessment. In January 2026 the Bureau of Industry and Security issued a rule permitting sales of the H200 and AMD's MI325X to China on a case-by-case basis rather than a presumption of denial. A presidential proclamation the following day imposed a 25 percent import tariff on both chips; exports were capped at 50 percent of each chip's total US sales; and exporters must certify the chips will not reach prohibited end users and submit them to third-party testing at US facilities before shipping. The Center for a New American Security estimates the cap allows China roughly 850,000 H200-equivalents, or close to 900,000 counting the AMD parts.

Read plainly, the United States moved a chip it had banned into a licensed, taxed and metered channel; Blackwell-generation parts stay out.

Digital Sovereignty: France, India, and the National Model Strategy

High-quality open weights have created a new policy option for a growing number of countries: building national AI capability on an open-source foundation rather than on US or Chinese commercial APIs.

That is sometimes called AI sovereign capability or strategic autonomy in AI: the principle that a nation's systems in government services, national security, healthcare and critical infrastructure should not depend wholly on foreign providers that can change pricing, restrict access or rewrite terms unilaterally.

France has been the most explicit and systematic about this strategy. Mistral AI — the French AI startup co-founded by former Google DeepMind and Meta researchers — has received substantial French state support and has made open-source releases a defining element of its identity. The French government has explicitly cited Mistral as a demonstration of European technological sovereignty. Emmanuel Macron's 2024 address on AI strategy prominently featured the open-source model as a vehicle for European independence from American and Chinese AI providers.

India is pursuing sovereignty at a different scale. The BharatGen initiative, discussed in the context of multilingual AI, is partly a language inclusion project and partly a sovereignty project: India's goal is to have AI foundation models developed by Indian institutions under Indian governance, available to Indian businesses and government agencies on terms that India controls. The explicit framing is that AI is infrastructure, and no nation should allow critical infrastructure to be foreign-owned and foreign-controlled.

Germany has invested through the LAION (Large-scale Artificial Intelligence Open Network) project, whose training datasets and open models are meant to free European research from US-controlled data pipelines. Germany's Federal Office for Information Security has published guidelines favouring open-source AI for government deployments on auditability and sovereignty grounds.

Nations that cannot compete in closed AI are finding open-source models a workable foundation for national capability, which gives agency to countries that would otherwise be pure technology importers.

Mistral and the European Open-Source Model

Mistral AI is the most sophisticated integration of technical ambition, commercial model and political positioning in open-source AI.

Founded in 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix — all former researchers at DeepMind and Meta AI Research — Mistral has released a series of models that have consistently punched above their size class. Mistral 7B, Mixtral 8x7B, Mistral NeMo, and Mistral Large have each introduced technical innovations that have influenced the broader field: the grouped-query attention, sliding window attention, and mixture-of-experts architectures pioneered in Mistral models have since been adopted in models from competing organisations.

Mistral's business model is unusual: it releases competitive open-weight models while selling premium services — fine-tuning, deployment infrastructure, enterprise support, access to its largest models — to commercial customers. The open releases are a deliberate strategy to build community and technical reputation, not a loss leader.

The model has attracted substantial investment. A €105 million seed round in June 2023, led by Lightspeed Venture Partners, was followed by a €385 million Series A in December 2023 led by Andreessen Horowitz and a €600 million Series B in June 2024 led by General Catalyst. The Series C closed on 9 September 2025 at €1.7 billion, at a post-money valuation of €11.7 billion, led not by a venture fund but by ASML, the Dutch lithography monopolist whose machines print the chips the models run on.

This investment brings together two technology leaders operating in the same value chain. — Arthur Mensch, CEO, Mistral AI

On 8 September 2026 Mistral raised a €3 billion Series D at a post-money valuation of more than €21 billion, announced under the explicit banner of making sovereign, open-weight AI the technology frontier. The investment case rests on the belief that the open-source ecosystem Mistral has built creates a durable commercial position — that the combination of open releases and commercial services is more sustainable than either pure open-source or pure proprietary approaches.

The Dual-Use Problem: When Openness Creates Risk

The case against open-source AI — or at least for some restrictions on the most capable open releases — deserves serious engagement: the risks are real and the governance gap is significant.

The core dual-use concern is that capabilities that enable beneficial applications also enable harmful ones, and that open-weight models, once released, cannot be recalled or access-controlled. A commercial API can implement safeguards, monitor for misuse, revoke access to bad actors, and cooperate with law enforcement. An open-weight model, downloaded millions of times and fine-tunable without restriction, cannot.

The biosecurity case is the most serious. Research published in 2024 by biosecurity experts at the Gryphon Scientific consulting firm and the Nuclear Threat Initiative found that large language models could provide meaningful "uplift" to individuals attempting to synthesise dangerous pathogens — that models with chemistry and biology knowledge could assist someone with partial biological knowledge to fill in gaps that would otherwise require specialised graduate training. The capability does not require frontier models — publicly available ones suffice — but it scales with capability, so open release of increasingly capable biological reasoning models raises the risk.

The cybersecurity case is less existential but more immediate. Capable open-weight code-generation models can be fine-tuned on malware datasets to produce offensive tooling that is operationally useful to actors who could not write it themselves, and a downloaded checkpoint cannot be revoked once that fine-tune exists.

The tension between the benefits of openness and the risks of dual-use capability is real and unresolved. Current approaches are ad hoc: Mistral and Meta publish acceptable use policies, but these are not technically enforceable for open-weight releases. The EU has since produced the first binding framework, and it carves openness out explicitly. The AI Act's general-purpose AI obligations applied from 2 August 2025, with providers who placed models on the market before that date given until 2 August 2027 to comply. Article 53 exempts from the technical-documentation and downstream-information duties any provider whose model is "released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available". The copyright policy and the training-content summary still apply.

The exemption has a ceiling. Article 51 presumes systemic risk for any general-purpose model whose cumulative training compute exceeds 10^25 floating-point operations, and a model designated as carrying systemic risk owes every obligation regardless of its licence. Brussels drew a compute line through the tension between openness and dual-use capability: openness buys relief below the line and nothing above it.

Who Wins the Three-Way Battle?

The reckoning has created a three-way competition for influence over AI:

Proprietary labs (OpenAI, Anthropic, Google DeepMind) are defending their positions by doubling down on post-training — the refinement, alignment and specialisation that adds much of the value users actually want — and on ecosystem lock-in: API integrations, developer tools, enterprise relationships and safety credentials that are hard to replicate with a downloaded checkpoint. Their moat is the full-stack service and the regulatory relationship, not the weights.

Open-source communities (Mistral, DeepSeek and Moonshot derivatives, the Llama ecosystem, Hugging Face's community of fine-tuners and evaluators) are benefiting from network effects: the more people use open models, the more fine-tunes, evaluations and applications flow back, improving capability at low marginal cost. The economic case for large enterprise buyers keeps getting stronger.

National governments are using open-weight models as instruments of sovereignty, building national capabilities that proprietary foundations would not allow. That puts capability outside the US and Chinese private-sector nodes, with consequences for global AI governance that are still unfolding.

The most likely outcome is not a decisive victory for any of these actors but a differentiated equilibrium: proprietary models dominating in frontier capability and safety-critical applications; open models dominating in cost-sensitive, privacy-requiring, and customisation-heavy applications; and national models serving sovereign government and infrastructure applications. The market is large enough for all three to coexist, but the distribution of power is shifting in a direction proprietary labs will not reverse.

The open-source reckoning is not the end of proprietary AI. It is the end of the assumption that proprietary AI is the only AI that matters.