Newsletter · July 12, 2026
Weekly Digest 28
China weighs curbs on overseas access to its top AI models, small firms swap Salesforce for custom apps built with coding agents, Meta's Muse Spark 1.1 shows its frontier rebuild closing the gap, and Nvidia pushes into the data-centre CPU market with Vera.
Topics we are tracking
China Begins to Protect Its Model Frontier
Source: Reuters — Beijing is looking at curbing overseas access to China's top AI models, sources say
Chinese authorities are considering restrictions on overseas access to the country's most advanced AI models. The Ministry of Commerce has held meetings with Alibaba, ByteDance and Z.ai over the past month, according to Reuters. The discussions covered closed and open-weight models, including systems that have not yet been released.
Chinese labs built much of their international position through open weights. DeepSeek, Qwen, GLM and Kimi could be downloaded, modified and served by foreign infrastructure providers, giving Western developers access to competitive models without sending their data to China or depending on a Chinese API.
This distribution generated influence and disrupted the market, perhaps much to the delight of the CCP, but it did not generate much direct revenue. When Fireworks, Together, AWS Bedrock or another American provider hosts an open-weight Chinese model, the inference provider collects the token revenue and the Chinese lab receives no direct payment. The benefit comes through adoption and external experimentation: open weights allow developers to deploy, modify and fine-tune the model across a wide range of real-world applications, producing public feedback, benchmarks and technical improvements that can inform future releases. This strategy is especially attractive for labs below the frontier because it accelerates adoption and makes their models a common base for further development.
Chinese models are increasingly important to Western startups and developers because they were capable and cheap. The growing use of Chinese models also gave Beijing a form of technological influence that its domestic cloud providers could not have achieved through direct distribution alone. Cursor, Airbnb and DoorDash have all worked with Chinese models that can be hosted independently and fine-tuned for specific tasks.
China is likely to push its labs towards a tiered release strategy similar to the one Alibaba already uses with Qwen. Alibaba offers its leading Qwen Max and Plus models through its own Model Studio APIs, while older and smaller Qwen models remain available as open weights. Alibaba's current model catalogue places Qwen3.7 Max and Plus at the top of its hosted offering, while its documentation separately supports deployment of open Qwen generations.
Keeping the strongest model behind an API would also improve the economics for the developer. Alibaba, DeepSeek or Z.ai could charge directly for access instead of allowing foreign hosting providers to capture the entire inference margin. We do not think many foreign customers would be allowed, or even willing, to rely on a Chinese-hosted API. The main attraction of Chinese models for Western companies has been the ability to run them on infrastructure they control.
We will publish a deeper analysis of the Chinese AI ecosystem next week, covering the roles of Alibaba, DeepSeek, ByteDance, Moonshot and Z.ai among others, and how Beijing is organising models, compute and distribution around national technological self-reliance.
A retreat by China from frontier open weights could also create more room for American open-model developers. OpenAI released GPT-OSS in August 2025, Google has continued to expand the Gemma family with Gemma 4, and Nvidia has built a broad open-source ecosystem around Nemotron, including model weights, datasets and training recipes.
Reflection AI is also trying to build a US-based open frontier lab, although it remains too early to judge its technical position. The company has attracted substantial capital and recently announced major government and infrastructure partnerships, including work connected to the US Genesis Mission and a reported compute agreement with SpaceX. It has generated a lot of attention but has not yet released a model.
Small Firms Are Building Their Way Out of Salesforce
Source: The Information — Small Firms Use Claude to Quit Salesforce
Small companies are beginning to replace Salesforce and HubSpot with custom applications built using Claude Code, Replit and Lovable. The Information found firms cutting software costs by 40% to 80% after replacing broad CRM subscriptions with narrower systems designed around their own workflows.
Many small teams use Salesforce, or CRM software more broadly, to hold customer details, track deals, log interactions and generate a limited set of reports. They rely on only a narrow part of the wider product while paying for a platform designed to support much larger and more complex organisations.
The SaaS application stack — CRM is the biggest slice
Coding agents are making those narrower applications easier to reproduce. A small company can now build a basic CRM around its own sales process without employing a large engineering team or hiring expensive implementation consultants. The customer database remains useful, but the surrounding application becomes easier to replace or route around.
This also puts pressure on seat-based pricing. SaaS pricing was built around humans logging into applications and interacting with them directly. As agents begin to read from, write to and act across these systems, the number of human users becomes less relevant to the amount of work being performed. A company may retain Salesforce as its system of record while reducing seats and moving more activity into external agents.
The cost of maintaining internal software should not be understated. AI has significantly reduced the cost of building the first version, but software still requires authentication, permissions, security, backups, data migrations, testing, documentation and integrations with email, billing and support systems. APIs change, business processes evolve and the person who built the original application may leave. This friction will get lower over time, but accountability and auditability are still open questions if this is entirely managed by agents.
Historically, the opportunity cost of building software in-house was often too high, and some argue that it will remain that way. Time spent maintaining an internal CRM is time that cannot be spent improving the company's core product or serving customers. Buying software transfers much of that burden to a specialist vendor, although implementation and integration still require substantial work.
We have long believed that AI will pressure the SaaS sector, though the effect will be uneven. Some software companies will benefit directly from more automated activity. More agents, applications and tool calls create additional systems that need to be monitored, secured and debugged. That expands the opportunity for companies in observability, security and other infrastructure-adjacent categories.
Software that mainly functions as a repository faces bigger obstacles. As agents become the primary interface, the application layer becomes easier to bypass, seat-based pricing becomes less defensible and more of the value shifts toward the underlying data, permissions and infrastructure.
Meta's Frontier Rebuild Starts to Show Results
Source: Meta AI — Introducing Muse Spark and the Meta Model API
Meta released Muse Spark 1.1 last week, three months after the first version of the model. The update improves performance across coding, tool use, computer control and multimodal tasks, while a new public API gives US developers direct access at pricing well below the leading models from Anthropic and OpenAI. Meta is now competing directly for external model workloads and aggressively pursuing the cost-performance frontier.
Muse Spark 1.1 does not place Meta at the top of the frontier yet. Developers who tested the model before release found it roughly comparable to Claude Opus 4.6 and GLM 5.2 for general agentic work, while still identifying weaknesses in coding workflows. We have not heard from any engineering teams planning to move meaningful token volume to the model, and Meta is not expected to reach parity with Anthropic or OpenAI until later this year, even under a bullish scenario.
Since rebuilding its AI organisation late last year, however, Meta has moved quickly. Llama 4 damaged Meta's position and prompted Zuckerberg to reorganise the company's frontier effort. Meta spent $14.3 billion on the Scale AI transaction and recruited researchers from OpenAI, Anthropic, Google and Thinking Machines. The first version of Muse Spark, released in April, still lagged contemporary Chinese open models. Spark 1.1 is a much more credible model only three months later.
Meta now has most of the inputs required to close the remaining gap: talent, compute and reinforcement-learning data. Data may be the most underappreciated asset. Meta has reportedly moved around 3,000 engineers into an applied AI organisation focused on creating reinforcement-learning tasks and environments. It can also draw on traces of real work across software engineering, advertising, finance and operations.
Its attempt to collect employee keystrokes and mouse movements for model training created significant controversy and has since been scaled back and paused. We nevertheless expect this type of work-trace collection to become common across large companies, although implementation will require clearer consent, stronger privacy controls and better data governance than Meta initially provided.
The company also has an aggressive compute ramp. SemiAnalysis projects that Meta will have more total AI compute than OpenAI or Anthropic by the end of 2026, although a meaningful portion will continue to support recommendations and advertising. Meta is simultaneously developing five data-centre campuses designed to exceed one gigawatt and building networking systems that can distribute workloads across multiple sites. Its advertising business can finance this buildout without requiring Meta to reserve much of the capacity for external cloud customers.
With Muse Spark, Meta has also abandoned its open-source-first strategy at the frontier. Its strongest model is closed-weight and sold through an API, allowing Meta to retain control over deployment and collect the inference revenue itself. The aggressive pricing suggests that Meta is prioritising adoption over near-term model margins. Its advertising cash flows allow it to undercut specialist labs while using the same models across consumer products, business messaging and developer tools. This is similar to Alphabet's playbook.
Meta still has ground to make up, and its frontier-AI agenda may clash with the demands of the wider product organisation. Alexandr Wang's team is focused on model progress, while Meta's existing leadership has to turn those models into products across advertising, social apps, messaging and hardware. Previous internal reporting pointed to tensions over ownership, credit and decision-making.
Nvidia Moves Further Into the CPU Market
Source: Reuters — Perplexity says it plans to use Nvidia's new CPU
Perplexity plans to adopt Nvidia's new Vera CPU after tests found it completed the company's agentic coding workloads around 1.5 times faster than the conventional processors it currently uses. Perplexity did not disclose the size of the deployment, but OpenAI, Anthropic and Oracle have also committed to using Vera. Nvidia expects the CPU to generate $20 billion in sales by the end of its current fiscal year.
The data-centre CPU market has historically been dominated by Intel and AMD, but Nvidia enters from a strong position because it already controls the most valuable component in an AI rack, much of the networking and interconnect layer, and the software used to program and operate the system.
Vera will be sold as part of larger systems such as the Vera Rubin platform, which gives Nvidia the advantage of designing the CPU alongside its GPUs, networking products and system software while optimising the entire roadmap around AI workloads. Nvidia has focused Vera on high single-threaded performance, memory bandwidth per core and predictable latency under full utilisation. The company says the chip completes a range of agentic and data-processing tasks up to 1.8 times faster than x86 processors.
Intel and AMD serve a much broader market, with processors that must support databases, virtual machines, enterprise applications and legacy software across a wide range of server configurations.
The balance between CPU and GPU demand is shifting as agents are able to do longer tasks. METR tracks the length of software tasks an agent can complete reliably and finds that frontier-model time horizons have roughly doubled every four to seven months. As these systems move from short coding requests towards workflows that take humans hours or days, a larger share of the workload consists of compiling code, running tests, operating browsers, querying databases and managing files and containers. These stages are often CPU-intensive and can leave the GPU waiting for the next model call.
This will increase the CPU capacity required for each unit of accelerator capacity. Longer-horizon agents spend more time acting in software environments between inference calls, making host-CPU performance and tighter system integration more valuable. The exact ratio will depend on the workload.
We have watched the CPU business closely as AI has become more agentic and believe Nvidia is best positioned in this market. Customers increasingly buy the CPU, GPU, networking and software as one integrated system, often through existing relationships with Nvidia and its server partners. Nvidia also benefits from its scale and long-standing relationship with TSMC, which manufactures both Vera and Rubin and already allocates substantial advanced-node and packaging capacity to Nvidia's roadmap. That supply-chain position should make it easier to ramp the full platform at volume and strengthens the case for Nvidia capturing more of the economics of each AI rack.
Seen on X
Other interesting stories
- Reuters: China's DeepSeek is developing its own AI chip, sources say
- Transformer Circuits: Verbalizable Representations Form a Global Workspace in Language Models
- Derek Thompson: Is This the End of Booze?
- Peter Harrell — Donald Trump: Trade Order Builder?
- xAI: Grok 4.5
- The Information: Khosla-backed startup claims breakthrough with largest-ever AI model on an iPhone
