In June, a Chinese lab released a frontier-grade language model, GLM-5.2 from Zhipu AI (still private), under a permissive open-source license with a one-million-token context window. A week earlier, Moonshot AI (private) shipped a trillion-parameter open model built for coding. Alibaba Group (BABA) saw its Qwen family cross one billion downloads on Hugging Face in January, passing Meta Platforms (META) and its Llama line as the most-downloaded open models on earth. Open weights have cracked the frontier open. As the model itself becomes a commodity, the question for investors now is — where does the value go next? Much of the answer sits inside the ROBO Global Artificial Intelligence Index (THNQ).Key Takeaways (as of June 2026):
Open-weight models now rival the closed frontier on coding tasks at a fraction of the cost. Alibaba’s Qwen alone passed one billion downloads in January 2026.
As models commoditize, the value moves to serving them. Nebius Group (NBIS) reports a customer cutting inference costs up to 26×. Meanwhile, Cloudflare (NET) now serves over 70 models at the edge.
Microsoft (MSFT) CEO Satya Nadella now frames the divide as “human capital and token capital.” THNQ’s 53 holdings span both the open-model makers and the infrastructure beneath them.
See more: AI Agent Infrastructure: Why Enterprises Need a Control Plane
The price of intelligence is falling fast. The bigger question is where those savings get redeployed. The inference, edge, and silicon sections below name the companies collecting them.Are Open-Source AI Models Closing the Gap With GPT & Claude?The gap between open and closed AI has narrowed to single digits on the benchmarks that advisors should care about — the ones that track real work. Open-weight systems from DeepSeek, Moonshot, and Zhipu (all private) now post coding scores within a few points of the best closed models. They also do it at a tenth to a thirtieth of the cost per token. Alibaba’s Qwen crossed one billion Hugging Face downloads in January 2026 and represents more than 50% of all open-model downloads globally. Meta’s Llama 4, released in 2025, remains open-weight and natively multimodal.
For most enterprise tasks, document analysis, customer triage, code review, and structured extraction, the open model is now the rational default on cost and data-privacy grounds alone. The frontier labs keep a real edge at the bleeding edge of novel capability, but that edge narrows with each quarterly release. The model, once the moat, is becoming the commodity input.How Distillation & Agent Swarms Make Open AI Models Production-ReadyTwo mechanisms are collapsing the capability gradient. The first is distillation. This is when a large “teacher” model generates high-quality reasoning traces, and a small “student” model trains on them, inheriting the capability without the compute. These can oftentimes be run on third-party clouds, or even self-owned hardware (including some consumer grade). The second is self-improving harnesses: a model proposes its own tasks, solves them, grades itself, and reinforces the better answers, compounding capability inside a company’s own walls with no data leaving the building.
Then those models get deployed in parallel. Agent fan-outs and swarms break a complex job into pieces and dispatch many specialized sub-agents at once, then merge the results. This is the expansion that matters: AI is moving past efficient problem-solving in the digital world toward learning how to solve, and actually being deployed, across multi-step, linked, and unpredictable real-world task sets. It is the same arc that we covered in Physical AI Goes Live, now playing out in software and on the factory floor at once.Satya Nadella on Human Capital & Token CapitalOn June 14, Satya Nadella published an essay arguing that a frontier without an ecosystem is not stable, and it reframed the whole debate. His line: “Every company is going to have to build what I think of as human capital and token capital.” Human capital is the judgment and pattern recognition of its people. Token capital is the AI capability that a firm builds and owns. His warning to staff was blunter: avoid “tokenmaxxing,” or routing every task through an expensive frontier model when a cheaper specialized one would do. Frontier AI for frontier work, and open, distilled models for the rest.
We made the same call from a different door. After the HumanX conference in March, we wrote that there are now two types of companies: the first treats headcount and agentic token workflows as equally important inputs into how the business runs, and the second is still in discovery, without a real AI strategy. The gap between them widens by the quarter. Nadella’s framing and ours converge on one point: the durable asset is the learning loop a company owns and keeps improving, while a rented model is only an input.Where AI Value Accrues: Inference, Edge & Custom SiliconFollow the compute. Inference, the work of running a model rather than training it, now consumes roughly two-thirds of all AI compute, up from one-third in 2023. When the model weight is free, the cost that remains is the serving of it, and that is where the margin lives.
Nebius Group (NBIS) built its Token Factory for exactly this moment: a managed platform serving more than 40 open models. One customer, the technology investor Prosus (PRX.AS), reported up to 26x cost reduction versus proprietary models. Another runs up to 200 billion tokens per day. Cloudflare (NET) routes inference across more than 70 models from data centers in hundreds of cities, placing the model next to the user, and acquired Replicate in April 2026 to deepen that catalog.
Underneath both, custom silicon is taking share: ASIC-based AI servers are forecast to reach 27.8% of AI server shipments in 2026, per TrendForce. MediaTek (2454.TW) projects more than $1 billion in AI ASIC revenue this year, Astera Labs (ALAB) grew first-quarter revenue 93% to $308 million on AI connectivity, and Credo Technology (CRDO) guided to roughly 120% revenue growth for its fiscal 2026. The weight trends toward zero. The network, the serving stack, and the silicon do not.Frequently Asked QuestionsAre open-source AI models actually competitive with closed models like GPT and Claude? On practical benchmarks, increasingly yes. Zhipu AI open-sourced GLM-5.2 on June 13, 2026 under an MIT license with a one-million-token context window. Early comparisons place it level with or ahead of leading closed models on math reasoning, though Zhipu published no official benchmarks at launch. Alibaba’s Qwen and DeepSeek’s latest models now score within single digits of the closed frontier on coding, at roughly a tenth to a thirtieth of the cost per token. Qwen alone passed one billion downloads in January 2026.
If AI models are becoming “free” or “too cheap to measure” when combined with increasingly smarter, more token-efficient setups and harnesses, where does the investment value go? To inference, delivery, and expansion of automation in the real world — including connectivity and robotics applications. Running models now consume about two-thirds of AI compute. Nebius (NBIS) reports up to 26x cost reductions on its open-model serving platform and Cloudflare (NET) serves more than 70 models from its edge network. Both are capturing recurring economics that the model layer is losing.
What is “token capital”? A term from Microsoft (MSFT) CEO Satya Nadella’s June 2026 essay. He examined the AI capability that a company builds and owns, set alongside its human capital. His point was that firms which only rent models risk having their expertise commoditized and firms that own a learning loop compound an advantage.What Open-Source AI Means for THNQ InvestorsThis is the rare structural shift where a thematic index captures both ends of the trade. The ROBO Global Artificial Intelligence Index (THNQ) holds the open-model makers giving their weights away — Meta, Alibaba, and Microsoft. It also holds the toll roads that those models run on — Nebius, Cloudflare, MediaTek, Astera Labs, and Credo. As intelligence commoditizes, the companies that we track on the infrastructure side collect the recurring economics that the model layer is shedding. For a primer on how that infrastructure layer fits together, see our Artificial Intelligence Content Hub.
Open-source is relocating the value in AI, moving it from the model to the systems that serve and deploy intelligence in the real world. Watch inference share of compute, ASIC server penetration, and enterprise agent adoption. Those three lines tell you where the token economy is heading, and THNQ is built to own the road it runs on.
Looking for regular updates? "Subscribe here":https://offer.vettafi.com/robo-index-subscribe for weekly insights on robotics, AI, and healthcare technology, delivered straight to your inbox.
For more news, information, and analysis, visit the Artificial Intelligence Content Hub.
THNQ is the underlying index for the ROBO Global Artificial Intelligence ETF (THNQ) and the L&G Artificial Intelligence UCITS ETF (AIAI.LN).
*VettaFi is the index provider for the funds referenced and receives an index licensing fee. The funds are not issued, sponsored, endorsed, or sold by VettaFi, and VettaFi has no obligation or liability in connection with their operation, marketing, or trading. This material is educational and analytical and is not individualized investment advice or a recommendation to buy or sell any security. Performance and data figures carry the as
