The Signal — April 18, 2026

DeepSeek is raising real money from real investors. OpenAI keeps losing senior leaders. Anthropic's CEO is in the White House. And Alibaba's latest open model runs on a fraction of its own parameters. Friday in AI.

DeepSeek Raises $300M at $10B Valuation — Its First Outside Money

DeepSeek, the Chinese AI lab that upended the industry with R1's efficiency gains in early 2025, is raising at least $300 million in its first external funding round at a $10 billion valuation. Until now, the company was entirely self-funded by its parent hedge fund, High-Flyer. The move signals that DeepSeek is building for the long haul, not resting on a single viral model.

For context, $10 billion is modest by current AI valuations — OpenAI sits at $300 billion, Anthropic above $60 billion. But that comparison misses the point. DeepSeek achieved competitive performance at a fraction of the training cost, which is exactly the kind of efficiency story that attracts smart capital. The fact that global investors are lining up despite escalating US-China tensions says something about how the market reads DeepSeek's technical position: too good to ignore for geopolitical reasons.

Stanford's 2026 AI Index, released this week, backs that read — China has nearly closed the gap with the US on model performance benchmarks. DeepSeek's fundraise is the market pricing in what the researchers already measured.

Sources: Reuters · Yahoo Finance · Brussels Morning

OpenAI Loses CPO and Sora Lead as the Side-Quest Purge Continues

Kevin Weil and Bill Peebles have left OpenAI. Weil was the company's former CPO and most recently led OpenAI for Science, an initiative that launched its Prism research platform. Peebles headed Sora, the video generation model that was reportedly burning through $1 million per day in compute costs. Both departures follow the shutdown of their respective projects — Sora's consumer app shuts down April 26, and the science team's Prism platform is being absorbed into other divisions.

This is the clearest signal yet that OpenAI is executing a real strategic pivot, not just trimming around the edges. The pattern from the last few weeks: shut down Sora (consumer video), dissolve the science initiative (research moonshot), beef up Codex (enterprise productivity), push the "super app" narrative (platform play). Everything that doesn't serve enterprise AI is getting cut. The company that once tried to do everything is now trying to do one thing very well.

The talent drain is also worth watching on its own. Weil and Peebles are senior departures stacked on top of an already long exit list. At some point, the question shifts from "who's leaving?" to "who's staying and why?"

Sources: TechCrunch · The Verge · NewsBytesApp

Anthropic's CEO Visits the White House — Mythos Forces a Thaw

Dario Amodei met with White House Chief of Staff Susie Wiles on Thursday, marking the first direct contact between Anthropic and the Trump administration since it branded the company "radical left, woke" and banned federal agencies from using its products. The thaw wasn't voluntary — it was forced by Mythos.

Multiple reports indicate that government agencies are increasingly anxious about being locked out of Anthropic's most capable model. Some officials have been quietly pushing for access despite the ban, creating internal pressure that eventually reached the top. When your political posturing crashes into the reality that a company you've banned has the best cybersecurity AI on the market, something has to give.

No deal was announced. But the meeting itself is the story. The administration that declared Anthropic a national security risk is now sitting across the table from its CEO. Any formal resolution — lifting the ban, granting selective access — would be a dramatic climbdown from the rhetoric of just weeks ago. Worth watching how they try to thread that needle without admitting they were wrong.

Sources: Politico · Politico (analysis) · NYT · Axios

Alibaba's Qwen3.6 Uses 3B of 35B Parameters, Beats Google's Gemma 4

Alibaba released Qwen3.6-35B-A3B, a mixture-of-experts model that activates only 3 billion of its 35 billion parameters per inference. Despite running on less than 10% of its own capacity, it leads Google's Gemma 4 across coding benchmarks: 73.4 vs 52.0 on SWE-bench Verified, 51.5 vs 42.9 on Terminal-Bench 2.0. It edges ahead on reasoning too (GPQA: 86.0 vs 84.3). Open weights, available on Hugging Face.

The mixture-of-experts approach isn't new, but these results are the strongest demonstration yet that you can get top-tier output while running a small fraction of your parameters at inference time. The cost implications are enormous. If you can match or beat a flagship model while using 90% less compute per query, the economics of deploying AI at scale look completely different. For anyone running models in production, this is the paper to read this week.

It's also another data point in the open-source resurgence story. Alibaba, Meta, Mistral, and others are consistently shipping open models that compete with or beat closed ones on specific benchmarks. The moat around proprietary weights keeps getting thinner.

Sources: The Decoder · Qwen Blog · Hugging Face

On the Editor's Desk

Today's four stories are loosely connected by a single question: who controls AI's future? DeepSeek's fundraise says China isn't ceding that fight. OpenAI's departures say the company is betting everything on enterprise dominance. The White House meeting says governments can't ignore capability even when they want to. And Qwen3.6 says open-source keeps eroding the advantages of keeping things closed.

Honorable mention: the Musk v. Altman $134B fraud trial, with jury selection set for April 27. WIRED published a deep preview this week. We'll cover it when proceedings actually begin — the pretrial noise-to-signal ratio is too high right now.

Dropped from consideration: the Stanford AI Index (already four days old and widely covered), and various AI regulation roundtables in Congress that produced more rhetoric than substance.