Diffusion CapitalDiffusion Capital
Back to insights

Newsletter · June 14, 2026

Weekly Digest 24

The U.S. government orders Anthropic to suspend foreign access to its frontier Fable and Mythos models, turning the model itself into an export control; China tightens its grip on indium phosphide, the optical substrate that wires AI clusters together; and Google second-sources TPU production across Intel, Samsung, and TSMC to route around the leading-edge bottleneck.

Topics we are tracking

Anthropic's Fable ban turns the model into the export control

Source: Anthropic — Mythos and Fable access update

Anthropic released Claude Fable 5 this week, the first generally available version of its Mythos-class models. Fable was supposed to be the compromise version: the same underlying model class as Mythos, but with safeguards around cybersecurity, biology, chemistry, and model distillation. Mythos 5 remained restricted to trusted cyberdefenders, infrastructure providers, and selected researchers.

We tested the model on our internal benchmarks and were genuinely impressed by its capabilities. The model had taste. It handled reasoning-heavy workflows well, and we were preparing to use it more broadly across our internal stack. Unfortunately, our excitement did not last long.

The U.S. government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals, citing national-security concerns. The order also applied to foreign-national Anthropic employees. Anthropic disabled both models for all customers while it works through compliance.

Until now, U.S. AI policy has focused mostly on the hardware layer: GPUs, semiconductor equipment, foundry access, and the supply chains needed to train frontier systems. The assumption was that controlling chips would control the frontier. The Fable order suggests that this is no longer enough and that the model itself is becoming the export control.

This creates a new dilemma for companies outside the United States, and even for domestic companies that depend on globally distributed teams. Most companies cannot build frontier models in-house. The cost is too high, and they do not have the talent pool. Even companies with strong technical teams are forced to choose between closed U.S. frontier models and open models that are cheaper, more customisable, and increasingly Chinese.

Open models have had strong adoption within the startup ecosystem for a while. They can be hosted locally, fine-tuned locally, and governed under domestic rules. But they also sit behind the frontier. Depending on the task and who you ask, they are roughly 6–9 months behind the best closed models. For many workflows that is acceptable. For the highest-value reasoning, coding, research, and agentic tasks, it often is not.

Our assessment remains that large frontier models trained at scale continue to outperform smaller specialised models on the most valuable general-purpose workflows. Domain customisation matters, but raw capability still matters more than many people want to admit. That increases the dependency problem for foreign companies. The better the frontier models become, the harder it is for companies to avoid relying on them.

Artificial Analysis Intelligence Index

64.9
61.4
60.2
57.2
56.6
55.3
54.7
53.9
53.8
53.2
52.2
51.7
Claude Fable 5with fallback
Claude Opus 4.8
GPT-5.5
Gemini 3.1 Pro
Qwen3.7 Max
Gemini 3.5 Flash
MiniMax-M3
Kimi K2.6
MiMo-v2.5-Pro
Grok 4.3
Muse Spark
Claude Sonnet 4.6
Artificial Analysis Intelligence Index v4.0, a composite of 10 evaluations — GDPval-AA, τ²-Bench Telecom, Terminal-Bench Hard, SciCode, AA-LCR, AA-Omniscience, IFBench, Humanity's Last Exam, GPQA Diamond, and CritPt. Top 12 of the index shown; Claude Fable 5 (with fallback) leads at 64.9, ahead of the next-best model by ~3.5 points. Source: Artificial Analysis.

We have asked ourselves this question before: what can you do when you do not own the intelligence?

Our answer has been to diversify the internal model stack. Not because every model is interchangeable, but because different models sit at different points on the frontier of capability, cost, latency, customisability, and reliability. Fable adds another dimension to that assessment: jurisdictional risk.

This strengthens the sovereign AI argument. Europe can use open-source models as a hedge against foreign platform dependency. But if frontier capabilities are increasingly treated as national-security assets, then powerful open models will also face more pressure. Open source is the best hedge against dependency, but it is also the hardest model to control once capabilities become dangerous enough.

China finds the photonics chokepoint

Source: Reuters — China's control over indium phosphide exports threatens AI data-centre rollout

Indium phosphide is used in the optical chips that move data through fibre instead of copper. That matters because AI clusters are increasingly constrained by compute and data movement. As clusters get larger, denser, and more power hungry, the interconnect layer becomes more important.

Optical links can move data faster, over longer distances, and with lower loss than copper. The larger the cluster, the more valuable that becomes.

China has tightened export licences for indium phosphide since February 2025. Reuters reported that China produced around 70% of global indium in 2024, while AXT and Sumitomo dominate global indium phosphide substrate production. Prices for six-inch indium phosphide wafers have reportedly risen around 250% to roughly $5,000.

The AI stack is full of small markets that suddenly matter. GPUs were obvious. HBM became obvious. Power is becoming obvious. Optical substrates follow the same pattern.

China does not need to block finished AI products to create pressure. It can slow the inputs. Gallium, germanium, graphite, rare earths, and now indium-related materials all fit the same pattern. Export licensing is more flexible than an embargo. Shipments can be delayed, approved selectively, or used as leverage in broader negotiations.

The Western response will be diversification, but new capacity is slow. Reuters notes that new plants can take two to three years to come online, and qualification cycles are long. For high-performance optical components, customers do not switch suppliers overnight.

The companies that matter here are the ones with supply and demand exposure. Sumitomo becomes more important as the main non-Chinese substrate alternative. AXT has the right exposure, but also the wrong geography, because much of its production sits inside the Chinese licensing regime. Coherent and Lumentum matter further down the stack, which is why Nvidia's direct investments into both are interesting. Marvell's acquisition of Celestial AI also looks more relevant in this context, because photonics is moving from a networking sub-theme into the AI infrastructure stack.

The other point is allocation. In a normal shortage, price does some of the work. In a strategic shortage, relationships do more of the work. Nvidia, the hyperscalers, and the largest optical suppliers will be first in line. Smaller buyers will get worse lead times, less certainty, and worse economics.

Google starts second-sourcing the TPU stack

Source: The Information — Google, Nvidia consider Intel as a backup chip manufacturer

Google is reportedly spreading future TPU production across Intel, Samsung, and TSMC. The Information reported that Google has ordered more than three million TPUs from Intel for delivery in 2028. It also reported that Google is in talks with Samsung to manufacture part of Icefish, its next-generation TPU.

Google is already working with MediaTek on Icefish, but TSMC is still expected to make the main compute die. Samsung would manufacture the memory-interface component on its 2nm process. Intel would become another manufacturing route for future TPU volume.

The TPU is becoming part of Google Cloud's AI infrastructure product. If Google wants to sell TPU capacity to external customers, it needs performance, volume, pricing visibility, and supply certainty.

TSMC remains the best manufacturer at the leading edge, but its capacity is also the most contested. Nvidia, AMD, Apple, Broadcom, Qualcomm, hyperscaler ASIC teams, and AI labs are all competing for the same advanced-node and advanced-packaging capacity. Google cannot leave the TPU roadmap exposed to one manufacturing bottleneck. Nvidia in particular has absorbed a large share of TSMC's AI-related capacity.

Google is building a broader industrial base around TPU production, with several suppliers each taking part of the stack. That should improve Google's leverage. It should also make TPU capacity more credible as a cloud product. If Google can secure enough volume, it can use TPUs to lower its own inference and training costs while offering customers an alternative to Nvidia-based instances.

Intel also fits into the broader reindustrialisation and onshoring agenda of the Trump administration, as well as the market's desire to hedge against a future Taiwan escalation. Expectations for Intel are extremely high, not just from Google but from the wider industry. The company has made the right strategic moves, but now has to deliver against very high expectations.

Seen on X

Other interesting stories