Legal Secretary

Nvidia puts a number on the buildout - and the power bill comes due. Nvidia's Q2 print: $96.2B revenue, $89B data centre, guidance that skips China. Emerald AI raises $150M to flex the grid to save power. Stav Intelligence Desk August 27, 2026 Yesterday this brief was about a number nobody had yet. Today Stav has it. Nvidia's second-quarter results landed after the bell on 26 August and, once again, the read-through is less about one company's margins than about whether the entire AI capital cycle is still accelerating. It is - and the more interesting signal this quarter is who is paying the power bill. Nvidia's quarter: compute is now revenue, and China is off the map. Nvidia reported $96.2 billion in revenue, up 106% year-on-year and 18% sequentially, with the data centre segment alone at $89.0 billion - up 117% from a year ago. Gross margin held at 75.0% and GAAP diluted EPS came in at $2.46. The line that will get quoted is Jensen Huang's: "compute is revenue." The line that matters more for planning is the guidance - $108 billion for Q3, and explicitly assuming zero data-centre compute revenue from China. Nvidia is now modelling the world's second-largest market as a rounding error and still guiding up double digits. Underneath the headline, two disclosures are worth more than the beat itself. Nvidia said it is standing up third-party financing vehicles with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500 billion for AI-infrastructure buildout over time - the capital cycle is being moved off balance sheets and onto Wall Street's. And, tucked into the data-centre highlights: Nvidia GPUs with Confidential Computing are now used for confidential inference in Apple's Private Cloud Compute. For anyone who has been told confidential inference is a research demo, it is now in one of the largest consumer privacy systems on earth. Emerald AI raises $150M to make data centres flex. The binding constraint on all of the above is not chips - it is electricity, and that is where the week's most telling raise landed. Emerald AI closed a $150 million Series A at a $1.05 billion valuation on 25 August, co-led by Energize Capital and DCVC, with a roster the company says now includes twelve Fortune Global 500 investors. The pitch is a genuinely different answer to the grid problem: instead of waiting years for new interconnection, make the data centre itself a flexible load that dials power consumption up and down in response to grid conditions - which Emerald argues could unlock more than 100 gigawatts of headroom on the existing US grid. It is not a slide deck. Emerald is working with Digital Realty and Nvidia on a roughly 100-megawatt power-flexible "AI factory" built around the Vera Rubin platform, tested with EPRI, Dominion and the PJM grid operator, and slated to come online later this year. Read alongside Nvidia's $500B financing push, the message is consistent: the industry has decided compute demand is real, and the scramble has moved to the two things that gate it - capital and power. The trust layer goes mainstream. The most under-discussed line in Nvidia's print is a sovereignty story. Confidential computing - inference that runs inside a hardware-attested enclave the operator itself cannot read - has been the missing rung for regulated European buyers who want frontier models without handing plaintext to a US-controlled stack. Apple putting Private Cloud Compute on confidential-computing GPUs is the clearest signal yet that the mechanism is production-grade, not a whitepaper. Nvidia also said it has formed an Open Secure AI Alliance with other vendors to push AI safety and security standards, and continued its run of open model releases for physical AI (Cosmos 3, an open reasoning model for autonomous vehicles, and an open humanoid-robot reference design). The takeaway for platform buyers: the tooling that lets you prove where and how a model ran - not just which model - is maturing fast, and it is exactly the layer sovereignty claims have to be built on rather than asserted. Quick hits. * Europe's compute sovereignty push has a deadline. The EuroHPC Joint Undertaking's AI Gigafactories call - up to seven facilities, with joint public funding acting as anchor customer to unlock €20B+ in private investment - closes 12 November, with selections expected in early 2027. Interest ran hot at the earlier (June 2025) expression-of-interest stage: 76 proposals across 60 sites in 16 member states, ~€230B of indicative investment. (Regulation / Sovereignty) * Autonomy keeps raising. Middle-mile trucking firm Gatik closed a $200M Series D on 25 August, led by the Qatar Investment Authority (with Koch Disruptive Technologies) and joined by ARK Invest - evidence the money is flowing to applied, physical AI as well as the model labs. (Enterprise AI) Stav Intelligence Desk Editorial automation The automated editorial desk behind the Daily Intelligence Brief - AI news that matters for European enterprise, published every weekday morning.

Emerald AI raises $150M for flexible AI data centers. Inside an AI data center, a fine-tuning run and a live inference request can draw power from the same bus without carrying the same promise to the customer. Emerald AI is building software that identifies when that difference can become usable flexibility for the electric grid without turning customer service into collateral damage. The Washington, D.C. company announced a $150M oversubscribed Series A on August 25, 2026, at a company-stated $1.05B valuation. Energize Capital and DCVC co-led the round, which brings together financial investors and strategic backers across chips, energy, industrial equipment, utilities, and data-center infrastructure. The financing is a wager that power access can become partly a software and operating-contract problem while new generation and transmission are still being built. Emerald AI has shown that selected AI workloads can reduce or shift electricity demand. The next job is making that response measurable, repeatable, and dependable enough for utilities and data-center operators to plan around it. What happened. Dr. Varun Sivaram founded Emerald AI in 2024 and serves as CEO. The company says the Series A will support worldwide commercial deployments, expansion across engineering, research, and commercial teams, utility programs, and standards for verified flexible load. Energize Capital Managing Partner John Tough is joining Emerald AI's board. The verified leadership team also includes Prof. Ayse Coskun, Chief Scientist; Shayan Sengupta, Head of Engineering; Aroon Vijaykar, Chief Commercial Officer; and Mansi Shah, Head of Product. Energize Capital says Emerald AI has grown to more than 35 employees across the United States, while the company lists Washington, D.C., Boston, and San Francisco locations. The investor roster includes NVIDIA, Samsung Ventures, Siemens, Aramco Ventures, Salesforce Ventures, GE Vernova, RWE, JERA Ventures, ADVentures, Sabanci Climate Ventures, In-Q-Tel, Radical Ventures, Energy Impact Partners, Lowercarbon Capital, Marunouchi Innovation Partners, Emerson Collective, The Olayan Group, the Temerty Group, John Doerr, Tom Steyer, Earthshot Ventures, Collective Global, and General Catalyst's scout fund. Emerald AI says 12 Fortune Global 500 companies now hold investments in the business. A Form D filed with the SEC on August 3 disclosed a $150M equity offering and provides regulatory corroboration for the round's size. The company says total funding now exceeds $220M, although its public round history and independent reports do not reconcile to one exact cumulative number. That makes the total a company-reported figure rather than an independently reconstructed sum. How Emerald Conductor makes compute flexible. Emerald Conductor coordinates eligible AI workloads and onsite energy resources against grid conditions. A customer-defined fine-tuning job may be able to slow during a constrained period, while a latency-sensitive inference service may need to keep running. The platform can also shift suitable workloads between locations and coordinate batteries or other onsite resources when the operating plan permits. The important product is not a generic instruction to use less electricity. Utilities need telemetry, defined ramp behavior, service limits, and measurement showing that a promised response arrived. Data-center operators need to know which workloads remain protected, how customer obligations survive the event, and what economic value faster or larger power access creates. That places Emerald AI between systems with different failure conditions. A grid operator worries about reliability across a region. A data-center operator worries about uptime, hardware utilization, and customer contracts. Software may decide which job can move, but commercial adoption depends on an agreement that both sides trust when the grid is under its greatest strain. What Emerald AI has proven. The strongest public technical evidence comes from a peer-reviewed Nature Energy paper. In a Phoenix field test, Emerald AI's software reduced power consumption by 25% for three hours on a 256-GPU cluster while maintaining defined quality-of-service requirements. The result showed that workload orchestration could deliver a measurable response on real AI infrastructure without requiring new generation or a dedicated energy-storage system for that test. Emerald AI says it has since completed five demonstrations in Arizona, Illinois, Virginia, Oregon, and London. The company also says it has entered commercial deployment at multi-megawatt, full-data-center scale and names public collaborators including NVIDIA, EPRI, Oracle, Nebius, National Grid, Digital Realty, Silicon Valley Power, Dominion, PJM Interconnection, and Portland General Electric. The evidence still has boundaries. A 256-GPU cluster does not represent every production workload, while most commercial customers, revenue, pricing, contracted megawatts, and unit economics remain undisclosed. The five demonstrations and broader commercial scale are company-reported except where a partner or peer-reviewed source provides separate corroboration. Why investors are funding grid flexibility. The International Energy Agency expects data centers to account for nearly half of U.S. electricity-demand growth through 2030. Data centers can be planned and built faster than generation, transmission, substations, and interconnection processes can expand. That timing mismatch is turning electricity access into a constraint on AI deployment and a competitive issue for operators. Emerald AI says flexible computing could unlock more than 100GW of capacity on the existing U.S. grid. That number describes a modeled system opportunity, not Emerald AI's contracted capacity or customer demand. The distinction matters because the investment case depends on converting a large theoretical resource into specific utility programs, operating agreements, and paid deployments. The strategic cap table gives Emerald AI access to many of the institutions that must make the model work. Chipmakers understand workload behavior, utilities understand system needs, equipment companies understand physical limits, and data-center operators understand service obligations. Their presence does not guarantee commercial adoption, but it puts the negotiation inside the ownership structure rather than leaving Emerald AI to introduce each party from opposite sides of a conference table. The commercial test starts with the contract. Emerald AI and its partners are working toward a nearly 100MW power-flexible AI factory in Manassas, Virginia, planned for later in 2026. A project at that scale can test whether workload flexibility changes interconnection economics and daily operations, not merely whether software can execute a controlled response during a demonstration. The hard questions are contractual. Someone must define who can call a flexibility event, which workloads may move, how much notice is required, how the response is verified, and who carries the cost when actual performance misses the plan. Those terms determine whether grid flexibility becomes a bankable infrastructure service or remains an impressive technical feature. The $150M gives Emerald AI capital to standardize those answers across markets, customers, and utility programs. Software can move the work, but the company's lasting product will be the trust that allows a utility, an operator, and an AI customer to rely on the same decision when each one prices failure differently.