Macron and Merz Agree on an Alternative to Palantir — Karp Speaks of 'Witchcraft'

A week in which a Chinese startup has reshuffled the transatlantic race: Moonshot's Kimi K3 is playing in the same league as GPT-5.6 and Claude Sonnet 5—free to download, with 2.8 trillion parameters and a price that puts the entire industry under pressure. While everyone is staring at benchmarks, the real battles have long since shifted elsewhere: to infrastructure, regulation, and the question of who will actually turn the time saved into value.

France and Germany plan to build a Palantir rival together

In a joint declaration following talks between Emmanuel Macron and Friedrich Merz, France and Germany have agreed to develop a European alternative to Palantir's military AI software. The declaration states the goal of a 'European sovereign digital backbone' that will cover data-centric security, artificial intelligence, and cloud solutions. France's Arcadia, an AI-powered command and control platform, is mentioned as a starting point, alongside unspecified 'comparable German solutions'.

Both countries had already pushed Palantir out of their intelligence services. France's DGSI announced in June it would replace Palantir with ChapsVision's ArgonOS, six months after renewing its contract with the US company. Germany's BfV also chose ChapsVision for the same role, and the Bundeswehr has completely excluded Palantir from the procurement of its defense cloud. A senior NATO commander recently stated, according to Politico, that there is no real European alternative to Palantir's Maven software, which the alliance uses to process battlefield data.

The declaration extends beyond software: France, Germany, and the UK are exploring cooperation on long-range weapons with a range of 2,500 kilometers, supported by the capabilities of ArianeGroup. The German-French tank program MGCS, intended to replace the Leopard 2 and Leclerc, is launching a research program on autonomous driving, sensors, and networking. The troubled FCAS fighter jet was conspicuously absent; instead, both countries want to create a 'European standard for collaborative air combat.' Palantir CEO Alex Karp had previously described Germany's rejection in an interview with Bild as 'talks about witchcraft.' → Techpresso

Synthszr Take: The question is valid, but the answer contains a design flaw in its very first sentence. 'Arcadia, alongside comparable German solutions' translates to: two starting platforms, two national champions, one common label. This is the exact pattern that has paralyzed FCAS and MGCS for years, now being repeated for the most sensitive software layer. A sovereign backbone thrives on a single data logic, a single security model, and a single interface for intelligence and command data. If you build it twice and 'harmonize' it later, you pay the incoherence tax in uniform: more output pointing in no common direction, plus the ongoing burden of keeping two stacks permanently interoperable. Palantir's real advantage is that Maven is a single, end-to-end platform, while Europe is currently planning a division of labor by national flags. The test will come with the first procurement budget: will it flow into a single platform with clear leadership, or into two national prestige projects conceded to each other as a compromise?

Xi Jinping Declares Open-Source AI a State Doctrine

Chinese leader Xi Jinping has declared the openness of AI models a guiding principle at the World Artificial Intelligence Conference in Shanghai. In a list of four observations, he placed the 'principle of openness' first and called for seizing this 'rare, historic opportunity' to promote open source, openness, and collaboration. At the same time, Xi demanded more regulation of the technology. According to Business Insider, many of the largest Chinese models, such as DeepSeek, are already available as open source. → StrictlyVC

Synthszr Take: Proclaiming openness as a state doctrine is a clever move when you're not guaranteed to win the race for the best closed model. China is turning its own position into a strategic advantage: if DeepSeek, Kimi, and Qwen land in developer stacks worldwide for free, Chinese architecture becomes the foundation on which others build. This is soft power via the toolchain, packaged as a gift to the global community. Zuckerberg followed the same reflex with Llama: capture the standard to suffocate the competition; now a state is turning it into foreign policy. The catch is on the US, whose walled gardens suddenly look like the more expensive, cumbersome path, while Chinese models spread freely.

Alibaba's Chip Software SAIL Goes Open Source, Attacks Nvidia's CUDA

Alibaba's chip design unit, T-Head, announced at the World AI Conference in Shanghai that it will release its previously proprietary software stack, SAIL, as open source. SAIL is the fundamental software architecture for the company's Zhenwu series of AI chips and has been freely available to international developers since the same day. T-Head is thus joining a broader push by Chinese chipmakers, which, according to the South China Morning Post, also includes Huawei and Moore Threads. Their common goal: to build open, collaborative software ecosystems as an alternative to Nvidia's CUDA toolkit. → www.scmp.com

Synthszr Take: Nvidia's real moat was CUDA. Nearly two decades of a software layer that every university, every lab, and every toolchain has dug into—a mental lock-in par excellence. Once you've built on CUDA, you don't just switch; rewriting costs months and nerves. Alibaba is attacking this very point by giving away SAIL: the fastest way to destroy a pay-to-play moat is to place the tool next to it for free. Economically, this is sound thinking, as abundance destroys value, and an open stack makes Nvidia's proprietary layer replaceable piece by piece. Whether SAIL is technically good enough to convince developers to move is another matter.

Meta Wants to Sell Anthropic Up to $10 Billion in Computing Power

Meta is in talks to lease AI computing capacity to Anthropic in a deal worth up to $10 billion over two years, according to the New York Times. This would allow Meta to monetize its massive data center expansion externally for the first time, while Anthropic struggles with scarce capacity. For Anthropic, compute is currently the bottleneck limiting its growth. For Meta, it would be a new business model alongside its own AI program centered on the Llama models. → StrictlyVC

Synthszr Take: With Llama, Zuckerberg wanted to set the industry standard and cut off the competition's oxygen to grow. Now he's selling the oxygen tank to that very competition. Meta is renting computing time to Anthropic, a company whose models have surpassed Llama in almost every ranking. This is a moment of silent admission: the data center expansion is in place, costs billions to operate, and its own models alone don't justify it. So capacity becomes the product, and suddenly it's smarter to earn money from the utilization of others' training runs than to wait for your own model's lead, which no longer exists.

Apple Sends 40 Legal Letters to Ex-Employees at OpenAI

Apple has sent legal letters to around 40 former employees who now work at OpenAI. According to the Financial Times, citing unnamed sources, the letters demand that the individuals secure relevant documents and communications and meet with Apple's lawyers. The move follows a lawsuit Apple filed last week in a California federal court against OpenAI for allegedly stealing trade secrets related to manufacturing processes and unreleased products. Named in the suit are OpenAI's Chief Hardware Officer Tang Tan, a former Apple vice president, and engineer Chang Liu, who allegedly kept a company laptop and downloaded confidential hardware files after leaving. Apple accuses OpenAI of pushing applicants to participate in 'show and tell' sessions with actual components while they were still employed at Apple. OpenAI told Reuters it has 'no interest in the trade secrets of other companies.' → StrictlyVC

Synthszr Take: In California, post-employment non-compete agreements have been unenforceable for decades. This is why Silicon Valley works at all: people switch jobs, take knowledge with them, and this very circulation drives innovation. When a corporation can't hold on to its ex-employees with a contract clause, it resorts to trade secret law, and 40 legal letters to individual movers are a clear signal to the entire talent market. The legal truth in the cases of Tang Tan or Chang Liu is one thing; the chilling effect on the other thousand hardware engineers at Apple is another. Anyone considering a move to OpenAI, Google, or a startup now has to factor in a potential subpoena, which makes mobility more expensive, all without a court ever passing judgment.

EU Forces Google: Android and Search Must Open Up to ChatGPT and Claude

The EU Commission issued two orders under the Digital Markets Act on Thursday, compelling Google to grant competing AI assistants and search engines the same access to Android and Google Search that, until now, only its own services have had. According to The Verge, Google must provide third-party assistants like ChatGPT, Claude, or Perplexity with comparable system and data access, allowing users to deeply integrate these tools into the operating system instead of Gemini, including responding to voice commands like 'Hey Google' and accessing device hardware. The second order concerns Google Search: competing search engines and AI chatbots are to receive access to search data that Google has historically kept for itself. Google has until January 2027 to start sharing search data and until July 2027 for the Android changes. → AI Secret

Synthszr Take: Gemini's strongest selling point was never the model, but its wiring into the operating system: the assistant that's already there when you say 'Hey Google,' that accesses your apps and sensors without you having to install anything. Brussels is now opening up this wiring via an API and making it a user choice. This means Google loses the exact lock-in point that turns an interchangeable chatbot into a default assistant. From a product strategy perspective, this means Gemini has to win on an Android with ChatGPT as the system assistant, not on an Android that doesn't let ChatGPT in at all. It will be interesting to see the reflex of citing data protection and security to keep the interfaces tight, because Apple has already played that same card when it held back Siri AI in Europe.

SpaceX in Talks with the Pentagon for Billions in AI Data Centers

SpaceX is negotiating with the U.S. Department of Defense to provide data center capacity worth several billion dollars, according to the Wall Street Journal. This would give the Pentagon access to large-scale computing capacity for its growing AI applications, such as for intelligence services and military use. A final contract has not yet been signed; the talks could also fail, according to the report. SpaceX already provides the U.S. military with rocket launches, satellite communications, and support for missile tracking. → Techpresso

Synthszr Take: What's particularly interesting here is the customer base. Over the years, SpaceX has built a position with the Pentagon that almost no cloud provider can replicate: launches, satellite communications, missile tracking, and now computing time on top. This is diversification from a position of power: the military is the anchor client, stacking multiple layers onto a single supplier relationship. For the Department of Defense, this is both convenient and risky, as each new contract increases dependence on a single provider who already holds too many critical strings. Amazon has set the precedent with $50 billion for government compute, but Amazon doesn't launch rockets. That's precisely the unique advantage Musk is leveraging: SpaceX is already up in orbit and selling the compute time along with it.

Netflix Used Generative AI in Around 300 Titles, and Barely Any Viewer Noticed

In its second-quarter report, Netflix disclosed that around 300 titles on the platform have used generative AI, predominantly in post-production. The company states it is increasingly using these tools to produce 'faster, cheaper, and at higher quality.' Netflix cites titles like The American Experiment, Glory, and Brasil 70: A Saga do Tri as examples, where AI was used for complex sequences such as enlarging crowds and historical mass scenes. Co-CEO Ted Sarandos said during the analyst call that the docuseries The American Experiment contains 17 minutes of 'AI-enhanced' footage, created twice as fast and at half the cost of previous methods. → MyClaw Newsletter

Synthszr Take: 300 titles, and the great wave of outrage never came. That's the real finding: the viewer notices nothing, and that's precisely why it works. The whole debate about AI in film thrives on the idea that you can distinguish a synthetic frame from a real one, while 17 minutes of enhanced crowd scenes simply play out like any other establishing shot. Streaming audiences judge an image by its effect, not its origin, and the effect is right. The proliferation is happening quietly because it hides in battle panoramas and crowds, where nobody is counting the faces anyway.

Moonshot AI Catches Up to the US Leaders with Kimi K3

On July 16, Chinese company Moonshot AI released Kimi K3, which it claims is the largest open AI model ever with 2.8 trillion parameters and a one-million-token context window. The model is available via web and API, and Moonshot says the weights will be released on July 27. In the provided benchmarks, K3 is on par with Anthropic's Opus 4.8 and OpenAI's GPT-5.5, trailing behind Claude Fable 5 and GPT-5.6 Sol, but it leads the Arena leaderboard for front-end web design with a score of 1679. On the Artificial Analysis Intelligence Index, the model scores 57 points and requires 21 percent fewer output tokens than its predecessor, K2.6.

Architecturally, according to AI Breakfast, Kimi K3 uses 896 experts, of which only 16 are activated per request, combined with Kimi Delta Attention for 6.3x faster decoding. In terms of cost, Moonshot charges $3 per million input tokens and $15 per million output tokens. For comparison: GPT-5.6 Sol costs $30 for output, and Claude Fable 5 around $50. The cost per task, according to Artificial Analysis, is $0.94, about half that of Opus 4.8.

The launch coincided with a speech by Xi Jinping, who described AI development as a 'symphony of international cooperation' and criticized US export restrictions. On the same day, according to Superhuman, 29 countries, including China, Russia, Serbia, Belarus, Cuba, Brazil, and Venezuela, founded the World AI Cooperation Organization based in Shanghai, without the participation of the US or Western Europe.

The markets reacted immediately: the Nasdaq fell by about one percent, according to the New York Times, as investors sold off chip stocks like Nvidia and Intel. Observers are classifying K3 as a 'DeepSeek moment,' but more serious this time: analysts recently estimated China's lag at six to nine months; Alberto Romero now puts it at zero. Moonshot is currently raising fresh capital at a valuation of $31.5 billion. In parallel, according to Dev.to, a price war is underway for coding models, with Claude Sonnet 5, GPT-5.6, and Kimi K3 appearing within a month and pushing token prices down to a quarter of previous levels. → AINews

Also: www.nytimes.com, newsletter@mail.synthszr.com, AI Breakfast, AI Valley, Morning Brew, AINews, Alberto Romero from The Algorithmic Bridge, AINews, Alberto Romero from The Algorithmic Bridge, Superhuman – Zain Kahn, Matthias von THE DECODER, AINews, AINews, AI Secret, Simon Willison from Simon Willison’s Newsletter, AINews

Synthszr Take: The entire US export control policy rested on a single assumption: cut China off from high-end compute, and the lag will remain at least six months. That assumption was nullified on July 16, and the Nasdaq priced it in within an hour. The bitter pill for Washington: the chip bans forced Moonshot to be efficient. 16 active experts out of 896 and 21 percent fewer tokens is exactly the kind of discipline you learn under scarcity. While Xi gathers 29 countries in Shanghai for an open AI alliance and slashes the price to a third of GPT-5.6 Sol's, the American industry is left with its expensive closed-source margin, which is now being eaten from below by a freely downloadable model from Beijing. It's no coincidence that, according to THE DECODER, the US military now prefers to act immediately with unfinished systems rather than wait for perfect ones. The real question for Washington now is what sanctions can even achieve when the opponent simply gives away the weights.