Google locks in 22-year nuclear deal; Oracle AI cloud backlog reaches $664B

Between September 4 and September 11, 2026, technology companies and model developers committed capital to physical energy supply, specialized vertical software, and independent compute infrastructure. Google signed a 22-year nuclear power agreement in Finland, Oracle reported $30 billion in new AI cloud contracts backed by customer prepayments, Mistral raised Europe's largest private funding round, OpenAI brought GPT-6 Astra into investment banking, and Anthropic documented how autonomous AI agents mutate malware in the field.

Google committed €13 billion ($15.1 billion) on September 9 to expand its AI infrastructure and clean energy portfolio in Finland over 2027 and 2028, supported by a 22-year nuclear power purchase agreement. The investment represents Google's largest capital commitment in Europe. 12

Google's arrangement with Finnish utility Fortum secures up to 50% of the electricity output from the Loviisa nuclear power station through 2050. Fortum said the contract provides the revenue certainty required to fund a lifetime extension and facility upgrades at the two-reactor plant. Google President and Chief Investment Officer Ruth Porat described the structure as "bring your own power," marking Google's first nuclear procurement agreement outside the United States. 2

The physical expansion centers on three new data centres in northern Finland, where winter temperatures fall below minus 10 degrees Celsius, allowing the facilities to use freezing ambient air for cooling. The plan adds a 94-megawatt commercial battery storage installation to absorb price swings and stabilize grid frequency during low-wind periods. Google also committed €31 million to regional education and training programs across Hamina, Kajaani, Muhos, and Vaala, including technical courses for 4,400 workers and data-centre apprenticeships for 100 students. 1

The Finnish government projects the initial buildout will support more than 37,000 construction jobs and contribute €3.6 billion annually to national GDP across 2027 and 2028. Once operational, the facilities will support approximately 7,000 permanent jobs. Alphabet previously raised its global capital expenditure target to between $195 billion and $205 billion for 2026, and the Finnish commitment demonstrates that hyperscale operators are treating utility-scale nuclear agreements as mandatory infrastructure requirements for future cluster sites. 12

Oracle booked more than $30 billion in new AI cloud contracts during its first fiscal quarter, lifting total revenue backlog to $664 billion. The company reported its financial results on September 10, beating consensus estimates of $639.89 billion compiled by Visible Alpha. 3

Customer prepayments shielded the company's operating cash balance. Capital expenditure reached $28.50 billion in the quarter, with $11.36 billion funded directly by advance customer payments. Chief Financial Officer Hilary Maxson stated that the majority of new capacity orders arrived through prepay terms or bring-your-own-hardware arrangements, enabling Oracle to preserve its full-year capital expenditure forecast of $90 billion to $95 billion while avoiding additional corporate debt. 3

Free cash flow recorded a deficit of $5.40 billion, beating analyst expectations of a $9.56 billion cash burn and narrowing from a $11.48 billion deficit in the third quarter of fiscal 2026. Total revenue grew 30% year over year to $19.30 billion, ahead of the $19.14 billion consensus, while adjusted earnings reached $1.92 per share against the $1.74 expected. Management raised its fiscal 2027 revenue outlook to at least $90 billion and increased its annual adjusted earnings target to $8.10 per share. 3

The conversion schedule provides visibility into physical server deliveries. Oracle expects approximately 50% of its $664 billion backlog to convert into recognized cloud revenue within 36 months. The commercial terms indicate that enterprise developers are accepting rigid prepayment schedules and multi-year lockups to guarantee compute availability, allowing cloud providers to expand physical capacity using customer liquidity. 3

Mistral raised €3 billion ($3.5 billion) in a Series D equity round announced on September 8, valuing the three-year-old French developer at more than €21 billion. The transaction represents the largest private equity funding round completed by a European technology company. 45

Samsung Electronics led the round, with Scaleup Europe Fund (managed by EQT) and existing backer PSG Equity participating as co-leads. New institutional investors included Advent, funds managed by BlackRock, and the Grand Duchy of Luxembourg. Previous investors ASML, Nvidia, Andreessen Horowitz, General Catalyst, Lightspeed, and Bpifrance contributed additional capital. Chief Financial Officer Johan Bergqvist confirmed that Microsoft, which signed a multi-billion-dollar European computing partnership with Mistral in July, chose to sit out the equity round. 45

Mistral is pacing toward $1 billion in annual recurring revenue by the end of 2026. The company operates across 20 countries and serves more than 125 enterprise deployments, including Airbus, ASML, and HSBC. Bergqvist stated that customer growth is accelerating in North America and Asia, with enterprises seeking alternatives to proprietary US cloud models following export restrictions that limited foreign access to frontier systems earlier this summer. 45

The company's commercial strategy frames model control across four dimensions: data that remains strictly within client perimeters, weights that can be tuned and hosted on private servers, predictable infrastructure pricing, and auditable production pipelines. By combining open model weights with custom enterprise compute, Mistral is capitalizing on European regulatory preferences for technical sovereignty, establishing a capital base capable of funding frontier training runs through 2027. 45

OpenAI launched ChatGPT for Financial Services on September 10, combining its GPT-6 Astra reasoning model with natively indexed market datasets from Daloopa, PitchBook, LSEG News, and Crunchbase. The vertical product is designed for investment banks, private equity firms, and equity research desks. 67

The system indexes and hosts partner data directly on OpenAI infrastructure. On the OfficeQA Pro evaluation—which measures an AI agent's ability to extract and interpret data across complex tables, charts, and footnotes in U.S. Treasury Bulletins—GPT-6 Astra scored 69.9%, improving on the 60.2% recorded by GPT-5.6 Sol. The interface allows financial analysts to click individual numbers in generated reports to inspect original spreadsheet cells, financial statement footnotes, and reconciliation adjustments, such as definitions of adjusted EBITDA. 6

OpenAI developed the offering alongside design partners Morgan Stanley and Evercore, focusing on company research, financial modeling, and client presentation decks. Administrators can upload firm-approved Excel, Word, and PowerPoint master templates to standardise output formats across advisory teams. The system includes shared authentication integrations with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva, and Moody's, allowing organizations to maintain existing commercial entitlements within the chat workspace. 6

Enterprise compliance mechanisms mirror strict financial services standards. Client corporate data is excluded from model training, sessions are encrypted in transit and at rest, and workspace retention schedules are configurable by administrators. Compliance officers can stream audit logs directly into internal surveillance systems through the OpenAI Compliance Platform, while administrators can separate user groups into discrete workspaces to maintain legal information barriers. The product shows OpenAI bundling vertical data feeds directly into its frontier model layer, competing with traditional desktop terminals for high-value knowledge workers. 67

Anthropic published its September 2026 Threat Intelligence Report on September 10, detailing how threat actors automated cyber operations, surveillance, and credential theft across seven harm categories between December 2025 and August 2026. The investigation confirmed that malicious activity occurred across Claude Haiku, Sonnet, and Opus models, while Claude Fable and Mythos safeguards blocked attempted abuse. 89

The report highlighted Russian state-nexus espionage group GTG-20006, linked to Midnight Blizzard and an operator using the moniker "JackPoterz." GTG-20006 built automated agent loops that monitored endpoint security alerts in victim networks. When commercial security tools detected deployed implants, the actor's AI agents automatically rewrote the malware's source code, compiled new binaries, and redeployed the mutated tools to maintain persistence without human intervention. 8

The group targeted more than 20 organizations across Europe and Ukraine, scanning remote access gateways across two dozen Ukrainian government departments. In one campaign, GTG-20006 exfiltrated mailboxes from two military drone component manufacturers, targeted an unannounced military drone maker, and reverse-engineered a proprietary drone computer-vision software development kit, recovering its hardware bill of materials and firmware design. The actor also breached three hotel Wi-Fi management vendors to execute DNS hijacking (known as CaptiveCrunch), serving malicious update prompts to visiting diplomats and defense contractors. In North Africa, the group compromised a government technology authority, stealing credentials that unlocked 300,000 national identity records and corporate registry entries for 500,000 companies. 8

Anthropic concluded that AI scaffolding has collapsed the tooling gap between state intelligence agencies and smaller operators. Public frameworks such as PentAGI allow single operators to orchestrate multi-stage reconnaissance, credential phishing, and data exfiltration at machine speed. The speed of AI-driven malware generation challenges signature-based endpoint detection, requiring security teams to deploy behavioral monitoring and runtime execution guards to counter automated adversary loops. 8

Cognition raised $2 billion in Series E funding at a $48 billion valuation on September 8. Led by Andreessen Horowitz and Accel, the round nearly doubled Cognition's $26 billion valuation from May. Cognition disclosed that its annualized run-rate revenue reached almost $900 million, up from $492 million four months earlier. The company also announced SWE-2, its latest coding model designed to run across multiple reasoning effort tiers. 1011

Positron AI raised $875 million at a $5 billion valuation on September 10 to commercialize its Asimov AI inference processors. The round was co-led by NEA, Atreides Management, and Silicon Graphics founder Jim Clark. Positron will use the capital to tape out its Asimov silicon, scheduled for production in late 2027, and build Titan multi-chip nodes engineered to serve 16-trillion-parameter models with context windows reaching 10 million tokens. Positron is currently deploying more than 50 racks of its first-generation Atlas inference hardware within Oracle Cloud Infrastructure. 12

Ant International, Mastercard, and Visa launched a joint "Know-Your-Agent" interoperability framework on September 10. Established under the Monetary Authority of Singapore's BuildFin.ai initiative, the agreement sets shared identity and risk-verification protocols for autonomous AI agents authorized to complete financial transactions. The framework links Visa's Trusted Agent Protocol, Mastercard Verifiable Intent, and Ant's Agentic Mobile Protocol, allowing merchants and payment rails to authenticate purchasing agents across banking boundaries. 13

Bank for International Settlements General Manager Pablo Hernández de Cos warned on September 10 that massive AI capital expenditure is creating financial stability risks. Speaking at the Global Fintech Fest in Mumbai, Hernández de Cos projected that the world's five largest technology firms will spend over $1 trillion on AI infrastructure across 2025 and 2026. He noted that capital spending is increasingly funded through debt and opaque private credit markets rather than operational cash flows, drawing historical parallels to speculative over-investment in the 19th-century railway mania and the dot-com era. 1415