Today's through-line is capital and infrastructure: DeepSeek reportedly closed about $15 billion in funding, while OpenAI is negotiating a raise of up to $30 billion ahead of a potential IPO. On the product side, Google introduced Agent Compute Engine for agents, Microsoft published its MAI-DxO diagnostic orchestrator in Science, and iFLYTEK launched Spark X2.6, trained on fully domestic compute, at its Global 1024 Developer Festival.

1. DeepSeek reportedly raises about $15 billion

According to Bloomberg, TechCrunch and Reuters, Chinese AI lab DeepSeek raised about $15 billion (roughly 100 billion yuan). More than 15 investors took part, including CATL, Tencent, Meituan, Geely, GigaDevice, Midea and NetEase, plus Boyu, CPE, Legend Capital and IDG; CATL and Tencent were among the largest backers. The round comes about two days after the US ordered federal civilian agencies to remove DeepSeek technology. In June the company raised more than $7 billion at about a $50 billion valuation; the new capital will help expand infrastructure and open models, and there has been earlier talk of a possible China IPO.

Why it matters: Amid tighter geopolitics and a global race on frontier models, DeepSeek is again using a mega-round to lock in compute and model investment.

Sources: TechCrunch; Bloomberg; Reuters via KSL

2. OpenAI works on up to $30B raise at $1.25T pre-money valuation

CNBC and TechCrunch report that OpenAI is working on a funding round of up to $30 billion at a $1.25 trillion pre-money valuation, led by UAE investors and Coatue. The raise is framed as a bridge ahead of a potential IPO, with a goal of signatures by the end of next week; the size is still fluid and may be paired with a UAE compute-capacity deal. OpenAI filed a confidential IPO in June, and its CFO has said it aims to go public in 2027. Its March post-money valuation was about $852 billion; earlier Bloomberg reporting had floated terms around $1.4 trillion. For context, Anthropic is seeking a $1.5–2 trillion IPO valuation and raised $65 billion at a $965 billion valuation in May.

Why it matters: Top-lab raise sizes keep climbing, and pre-IPO valuation fights are increasingly tied to compute deals.

Sources: CNBC; TechCrunch

3. Google launches Agent Compute Engine and expands agent governance

At AGNTCon+MCPCon, Google Cloud announced Agent Compute Engine: purpose-built compute for agents that already powers Gemini agent work tasks, with a preview coming in the next few months. Each agent runs in a hardware-isolated microVM (Cloud Hypervisor + Titanium), with rapid boot, copy-on-write memory sharing, and pause/resume when idle; state can persist on Hyperdisk. The open-source Agent Substrate on GKE claims up to 10x density and under-500ms suspend/resume in a test with about 200,000 agents. Google is also expanding agent governance across Cloud runtimes—Gateway, Registry, Identity and observability.

Why it matters: Agents are moving from “can run” to “can run at scale under governance,” with compute and control shipped together.

Source: Google Cloud Blog

4. Microsoft's MAI-DxO appears in Science: orchestrating multi-step diagnosis

Research on Microsoft's AI Diagnostic Orchestrator (MAI-DxO) and the Sequential Diagnosis Benchmark (SDBench) was published in Science on October 22. Starting from a brief presentation, the system simulates a physician panel that iteratively asks questions and orders tests. On complex NEJM cases it reached about 83% accuracy versus about 36% for physicians with references and about 22% without; on MIMIC-IV-ED emergency cases it reached about 80% versus about 47%. In cost-constrained setups it can also cut workup spend substantially. The authors stress this is a controlled text-only benchmark, not proof of superiority over specialists or real practice; they plan to open-source parts of SDBench and run real-world evaluation with Beth Israel Deaconess.

Why it matters: Medical AI is shifting from single-shot answers to multi-turn diagnostic orchestration—while real-world validation remains essential.

Sources: Microsoft AI; Science

5. iFLYTEK launches Spark X2.6: an edge–cloud model on fully domestic compute

At the 2026 Global 1024 Developer Festival, iFLYTEK released Spark X2.6, an edge–cloud large model trained on fully domestic compute, with stronger coding, agent and complex-task capabilities. It reports roughly 4x higher long-context training efficiency and about 4.5x higher long-horizon reasoning efficiency; a lightweight on-device Spark X2.6-4B is also shipping, with claimed 96.1% edge–cloud routing accuracy, alongside apps such as WinWork digital employees.

Why it matters: Edge–cloud models on domestic compute stacks keep moving into products, with emphasis on coding, agents and long-horizon tasks.

Sources: 21st Century Business Herald; Sina Finance; Phoenix Finance

---

In one line: Top US and Chinese labs keep trading mega-rounds for compute and runway, while agent infrastructure and medical diagnostic orchestration move toward scalable, verifiable next steps.