Enterprise Software Deal Wave: Tencent Leases 100,000 Oracle AI Chips for $7bn

The chips are not available inside China and sit in several Oracle data centres in South-East Asia, according to the Financial Times report, which was summarised by The Next Web on 2 October. Tencent is paying about 30% upfront, the FT's Zijing Wu reported. TNW said it had not independently verified the figures.

It is a lease, not an acquisition, and that distinction matters. Tencent is buying access to compute rather than owning it, which is cheaper on the balance sheet and easier to walk away from if US export rules tighten. Those rules currently allow overseas cloud leases of this kind, the FT reported. The timing is not accidental. Tencent's cash flow turned negative in the second quarter, at RMB 13.8bn ($2.06bn), the first negative figure in more than a decade, according to the FT. Chief financial officer John Lo attributed it to heavy AI infrastructure spending and prepayments for compute, DatacenterDynamics reported. Executives said the compute would go first to training bigger Hunyuan models, then to running them, and later Tencent Cloud could rent the capacity out to other customers. The company also backs Chinese chipmaker Enflame, which nearly tripled in its Shanghai debut in September, a sign that the domestic chip route is still being kept warm alongside the imported one.

What Tencent gets, and what it gives up

Tencent president Martin Lau has defended the spending in blunt terms: "We're comfortable in making significant investments in AI because not only is there a substantial upside potential, there is also clear downside protection." That quote was carried by TNW in its write-up of the FT report. Whether the downside protection holds depends on export policy staying where it is for five years, which is not something Tencent controls.

Chinese tech groups have rented compute abroad before. ByteDance brought in nearly 75% of Nscale's 2025 sales, the UK cloud firm's filings showed. The pattern is established: buy the brains where the rules allow, build the products at home.

Oracle's side of the deal is less flattering. Shares barely moved on Thursday after the report, Investor's Business Daily reported. They had fallen 29.5% this year going into the day. Oracle's free cash flow has been below zero for several quarters, IBD noted, because of its AI spending, and the company is spending on data centres more than it earns in a quarter. A $7bn lease with 30% upfront is useful cash for a company in that position.

The robot brain round next door

The same day, a very different kind of deal surfaced. FieldAI, a Californian software company that makes no robots at all, is set to raise $700M at a $10B valuation, according to The Next Web. That is five times the $2B mark set just over a year ago, according to a person familiar with the deal, as Business Insider reported. A term sheet is signed and the round has not closed. FieldAI did not respond to a request for comment.

FieldAI sells what it calls a universal general-purpose brain, meant to run humanoids, robot dogs, drones and industrial rovers. Founded in 2023, it sells to construction firms, data centre operators and defence companies. Revenue and customer contracts have passed $135M across more than 30 customers, up at least $35M since June. Skild AI, which is valued higher, reported $100M in run-rate revenue last month, on a stricter measure. The geography is the story. Europe's biggest robotics round this year built robots: NEURA raised up to $1.4B in June at around $7B, backed by Tether, Nvidia, Amazon, Bosch and the European Investment Bank. London's Humanoid raised $152M in July at $1.35B, with Bosch contracted to build its wheeled robots. The nearest European answer to FieldAI is smaller and less settled: Genesis AI, which splits about 60 staff between Paris, London and the Bay Area, was reported in July to be raising $500M for robot brains at roughly $3B. That one was a report. FieldAI has a signed term sheet, and the two valuations sit three times apart.

Three American brain-layer companies, Physical Intelligence near $11B, Skild AI at $14B and FieldAI at $10B, carry roughly $35B between them. Europe's comparable names are a fraction of that. Munich's Agile Robots was in early talks for $800M, and it builds the software as well as the machines.

One regulatory date matters here. Europe's Machinery Regulation starts on 20 January 2027 and covers machinery with self-evolving behaviour, which is what this software is sold to produce. No model of this kind has been assessed under the European rules yet, because they do not start until 2027.

Inside the enterprise: Forge, IAM and the agentic pitch

On 30 September, CoreWeave unveiled CoreWeave Forge, an integrated software platform for organisations to develop, run and continuously improve AI models and agents, announced at the company's Fully Connected event in San Francisco. Forge is available today in Free, Pro and Enterprise editions and unifies existing products with new software into a stack for training, inference, evaluation, observability and agent development, Data Center Knowledge reported.

"We have moved from a focus on being the best provider for AI-centric infrastructure to delivering a plethora of AI services to enable customers to build AI applications on our platform," said Corey Sanders, CoreWeave's senior vice president of product management, at a media briefing during the event. Early partners include VAST Data, CrowdStrike and ClickHouse. CoreWeave is in early phases of making Nvidia's Vera CPU for AI agents available, with some early access underway, and said the Nvidia Vera Rubin NVL72 is now available on its cloud platform. AI startup Cognition is the first customer to run workloads on the AI hardware platform and was up and running in two days, Sanders said.

IDC analyst Dave McCarthy read the announcement as a positioning move: "CoreWeave has recognized that it's great to get these big AI labs, but if they're going to build a long-term sustainable business, then they need to be able to appeal to a wider audience, including enterprises. That requires a bigger ecosystem of software."

The governance problem shows up in a separate framework paper. Identity and access management for AI agents treats each agent as a non-human identity with a human owner, a defined purpose, scoped authorization, an expiration and continuous monitoring, according to a guide published by The Hacker News on 28 September. The same piece names OWASP's Top 10 for Large Language Model Applications, which calls excessive agency LLM06: an agent granted broad functionality, permissions or autonomy exercises capability beyond its approved task. Configuration findings describe possibility; telemetry describes what occurred, the guide argues.

Buyers are already being told that the integration work, not the model, is the hard part. An MIT Technology Review custom report published on 2 October puts global AI investment at $2.5 trillion in 2026, up 44% from the previous year, and says the majority of enterprises are still not growing revenue through AI or fundamentally rethinking how they operate. The report is produced by the publication's Insights arm, not its editorial staff. Its core claim is that process redesign has to precede model selection. A paper submitted to arXiv on 30 September by Terry Dorsey and Kevin Huggins makes a narrower version of the same argument. Enterprise Representation Simplification is defined as reducing unnecessary representational complexity while preserving required information within a defined scope. Reductions in task-level complexity reduce the representational extent an AI system must identify, relate and interpret, the authors write, and text-to-SQL research provides evidence that reduced schema and reasoning complexity can improve reasoning accuracy.

None of this is settled. What the week shows is money moving in two directions at once: into compute that has to be leased because it cannot be owned locally, and into software layers that sit above the machines. The deal that closes first is not necessarily the one that lasts.