🇺🇸QumulusAI (Nasdaq: QMLS) reported $6.7 million in quarterly revenue, up 118%, with AI compute contributing 84%. #
QumulusAI reported second-quarter 2026 revenue of $6.7 million, an increase of $3.6 million, or 118%, from $3.1 million in the corresponding 2025 quarter. Compute-power revenue reached $5.6 million, while AI compute represented 84% of total revenue, up from 61% in the first quarter of 2026 and 43% in the second quarter of 2025.
The company reported a 67% gross margin and said GPU activations allowed revenue to grow faster than colocation costs. Deferred revenue increased by $30.5 million during the first half of 2026, while operating cash flow totaled $22.3 million.
QumulusAI expanded its deployed graphics processing unit fleet from 952 to 3,088 during the quarter, an increase of approximately 224%. It ended the quarter with 8 megawatts of high-performance computing capacity covered by executed lease and colocation agreements.
This was QumulusAI’s first quarterly report after its shares began trading on the Nasdaq Global Market under QMLS on July 16, 2026. The business traces its origins to data-center operator SPRE and blockchain managed-services provider WAHA, later adding GPU-as-a-Service operations through its acquisition of The Cloud Minders.
🇺🇸QumulusAI disclosed $282.5 million in cumulative signed contract value and purchased 1,632 NVIDIA Blackwell B300 GPUs. #
QumulusAI disclosed that cumulative signed customer contract value had reached $282.5 million. This figure represents signed contract value rather than recognized revenue, and the filing does not equate the full amount with revenue already earned or collected.
The company also reported signing more than $120 million of new customer agreements, including one three-year agreement valued at more than $71 million. It separately disclosed a GPU-as-a-Service agreement with diversified trading firm DRW.
QumulusAI said it purchased 1,632 NVIDIA Blackwell B300 graphics processing units to serve customer demand. It also entered a metropolitan Atlanta colocation agreement for up to 3.75 megawatts, with a right of first offer covering as much as 7 megawatts of additional capacity at the same site.
The filing identifies material concentration and deployment risks. A substantial portion of hosting revenue comes from a limited number of customers, much of current GPU-as-a-Service revenue is generated through one channel partner, and operations depend on access to equipment, power, data-center capacity and additional capital.
🇺🇸The National Science Foundation (NSF) awarded more than $290 million across eight Quantum Leap Challenge Institutes. #
The U.S. National Science Foundation said eight Quantum Leap Challenge Institutes will collectively receive more than $290 million. The institutes are intended to address scientific hurdles associated with developing quantum-based technologies.
A Yale-led multidisciplinary team received $37.5 million for a center focused on practical quantum error correction. Its work will address the design of practical and self-correcting quantum computers across the technology stack and provide approaches that industry could use to build more reliable systems.
Princeton University will lead another of the eight institutes with a $27.9 million award. The Princeton-led program will develop quantum-processor fabrication techniques and support education and workforce-training programs in quantum science and engineering.
The disclosed awards are research grants to institutions, not purchases of commercial quantum-computing capacity. The program’s scope includes quantum-computing hardware, manufacturing and error correction rather than conventional GPU or central processing unit infrastructure.
- National Science Foundation (NSF)
🇪🇸Spain prepared draft data-center rules covering water, energy, cybersecurity, EU establishment, and EU-localized data and metadata. #
Spanish government sources said a draft decree would impose stricter water-use, energy and cybersecurity requirements on data centers. The draft was scheduled to open for public consultation during the week, meaning the requirements were proposed rules rather than final regulations.
Under the draft, data-center operators would have to be established in the European Union. Data and metadata processed by the facilities would also have to remain within the EU, with mechanisms required to control access from third countries.
The proposal would apply stricter requirements to data centers handling public-sector information or data involving national security. Government sources described sustainability, cybersecurity and operating efficiency as criteria for selecting among the large number of proposed projects.
The source characterized Spain as a major European data-center destination because of its renewable-energy resources and available land. It also said the country had attracted artificial-intelligence infrastructure proposals worth tens of billions of euros, without identifying individual projects or committed deployment totals.
🇺🇸The U.S. Treasury launched its Quantum-Readiness Task Force to strengthen cryptographic protections for sensitive data and critical infrastructure. #
The U.S. Department of the Treasury announced the Quantum-Readiness Task Force following Executive Order 14412. The executive order established guidelines for strengthening cryptographic protections around sensitive U.S. data, critical infrastructure and the digital economy.
The task force is part of Treasury’s work to prepare cryptographic systems for quantum-computing risks. The department connected the initiative with securing U.S. innovation, strengthening supply chains and maintaining U.S. leadership in relevant technologies.
The announcement concerns security planning and cryptographic readiness rather than funding for quantum computers or the procurement of quantum-processing capacity. No task-force budget, hardware order or deployment capacity was specified in the supplied source.
- U.S. Treasury
- Quantum-Readiness Task Force
OpenAI plans to deploy its Jalapeño custom accelerator by year-end after reporting better response speed and AI work-per-watt than GB300 in tests. #
OpenAI said it plans to begin deploying Jalapeño in its compute infrastructure by the end of 2026. The custom accelerator was developed with Broadcom and is intended to support OpenAI’s artificial-intelligence models.
OpenAI chip lead Richard Ho said Jalapeño led NVIDIA’s GB300 comparison system in two tested areas: AI workload processed per unit of power and response speed. OpenAI described GB300 as the leading publicly benchmarked system used for the comparison; the supplied sources do not provide the underlying test configuration or numerical benchmark results.
The company said Jalapeño combines higher throughput with lower latency without sacrificing energy efficiency. OpenAI linked those characteristics to faster ChatGPT responses, more responsive Codex sessions and agents, and service availability as demand grows.
OpenAI will determine which of its models run on the accelerator. The company said customers could benefit from choices oriented toward either lower cost or higher performance, but it did not disclose deployment volume, manufacturing quantity or infrastructure capacity.
Google detailed its eighth-generation custom accelerators: TPU 8t for training and TPU 8i for inference. #
Google presented its eighth-generation Tensor Processing Unit family at Hot Chips 2026. Tensor Processing Units are Google-designed accelerators for machine-learning workloads rather than general-purpose central processing units.
The family separates its primary workload roles: TPU 8t is designed for model training, while TPU 8i is designed for inference. Training covers the computation used to develop or update models, whereas inference runs trained models to produce outputs.
Google is one of the hyperscale cloud operators developing its own training hardware. The supplied source identifies the two accelerator variants but does not provide memory capacity, power consumption, benchmark results, system scale or general-availability timing.
Microsoft presented Maia 200, its second-generation server accelerator designed specifically for artificial-intelligence inference workloads. #
Microsoft provided new technical detail on Maia 200 at Hot Chips 2026. Maia 200 is the company’s second-generation server-class artificial-intelligence accelerator.
The processor is designed specifically for inference, the stage in which a trained model processes inputs and produces responses. It is therefore an accelerator rather than a general-purpose server central processing unit or a complete server, rack or cluster.
The supplied source does not disclose Maia 200’s memory capacity, manufacturing process, power envelope, benchmark performance, deployment volume or availability schedule. It likewise does not identify a specific customer deployment or committed data-center capacity.
Intel’s Crescent Island AI GPU supports between 160 GB and 480 GB of LPDDR5X memory, emphasizing memory capacity. #
Intel disclosed additional details about its Crescent Island artificial-intelligence graphics processing unit at Hot Chips 2026. The accelerator supports configurations ranging from 160 gigabytes to 480 gigabytes of LPDDR5X memory.
The disclosed range refers to memory attached to an individual GPU configuration, not aggregate server, rack or cluster memory. LPDDR5X is the specified memory technology; the source does not identify high-bandwidth memory as the device’s memory type.
Intel’s presentation emphasized memory capacity as a central design characteristic. The supplied source does not provide computational-throughput benchmarks, power figures, pricing, shipment volumes or a commercial availability date.
NVIDIA Jetson Orin Nano 2 doubles its predecessor’s inference performance, reduces equivalent-performance power by 40%, and launches during the first half of 2027. #
NVIDIA announced Jetson Orin Nano 2 as an entry-level edge-AI and robotics computer. It is intended for local inference in robots, delivery and inspection drones, and visual-AI systems rather than data-center training clusters.
The new system delivers twice the inference performance of its predecessor within the same physical form factor. NVIDIA also said it consumes 40% less power when operating at equivalent performance.
NVIDIA reported that more than 3 million developers build on its robotics technology stack. Cognex, Doosan Bobcat and home-cleaning robotics company Matic were identified as early companies adopting or evaluating Jetson Orin Nano 2.
Ecosystem partners are developing carrier boards, hardware systems and reference designs around the product. NVIDIA said Jetson Orin Nano 2 modules and kits are expected to launch in the first half of 2027.
Digital Realty received a 50-megawatt allocation for a fourth Singapore data center, potentially lifting local capacity to 134 megawatts. #
Digital Realty said it received a 50-megawatt data-center allocation for a proposed fourth facility in Singapore. The company plans to locate the project on Jurong Island.
Digital Realty’s existing Singapore data-center capacity was reported at approximately 84 megawatts. If the new project becomes operational at the full allocated 50-megawatt scale, total local capacity would rise to approximately 134 megawatts.
The increase would be about 60% relative to the existing 84-megawatt base. The source describes the 50 megawatts as an allocation and proposed development capacity, not capacity that is already operating.
No construction cost, commissioning date, customer commitment, server count, rack count or accelerator inventory was provided. The figures describe facility power capacity rather than computing performance.
New Jersey seeks 150 megawatts of behind-the-meter storage, offering maximum annual incentives of $200 per kilowatt for 10 years. #
New Jersey is seeking 150 megawatts of behind-the-meter energy storage to support a virtual power plant. Behind-the-meter systems are located on the customer side of the utility meter, while a virtual power plant coordinates distributed energy resources.
The program offers maximum annual incentives of $200 per kilowatt. The incentive period is 10 years, preserving separate units for power capacity—kilowatts or megawatts—and program duration.
State regulators said the comparatively modest maximum reflects the “private resilience value of residential energy storage systems.” The source does not state that all 150 megawatts have been contracted or installed.
No individual project sizes, battery-energy capacity in megawatt-hours, award recipients or installation schedule were provided. The 150-megawatt figure is therefore a procurement target for power capacity, not a confirmed operating fleet.
Victoria opened Australia’s first offshore-wind auction for 2 gigawatts; bids close in August 2027, with contracts due later that year. #
The government of Victoria opened a request for proposals for 2 gigawatts of offshore-wind projects. The process was described as Australia’s first offshore-wind auction.
The state said the planned projects could generate enough electricity to supply approximately 1.5 million homes. That household-equivalent figure is a government estimate associated with the full 2-gigawatt program.
Proposals are due in August 2027, and the related contracts are expected to be finalized during the second half of 2027. The opening of the auction does not mean the capacity has already been awarded, financed, constructed or connected to the grid.
The source does not identify winning developers, individual project capacities, contract prices or commercial-operation dates. The 2-gigawatt figure represents the auction’s requested project scale.
QumulusAI signed a three-year, $99.6 million subscription for 128 bare-metal clustered NVIDIA B300 GPU nodes, requiring $16.9 million upfront. #
QumulusAI entered the customer contract on May 11, 2026. The customer subscribed to 128 bare-metal clustered NVIDIA B300 GPU nodes for a three-year term, with aggregate contracted fees of approximately $99.6 million.
The customer is required to make an upfront payment of approximately $16.9 million. The remaining fees are payable over the term of the agreement, so the full $99.6 million was not described as an immediate cash receipt or already recognized revenue.
The contracted unit is a GPU node, not an individual GPU. The filing does not specify how many NVIDIA B300 accelerators are installed in each node, and the 128-node count therefore should not be presented as 128 GPUs.
The agreement was disclosed in QumulusAI’s Form 10-Q as an executed customer contract. The filing also identifies broader risks involving equipment supply, power availability, facility interruptions, customer concentration and dependence on a single GPU-as-a-Service channel partner.
NVIDIA Dynamo’s Shadow Engine Recovery restores large-language-model inference capacity within seconds following an engine-process failure. #
NVIDIA introduced Shadow Engine Recovery for NVIDIA Dynamo, its software framework for large-language-model inference serving. The feature is designed to restore serving capacity within seconds after an inference-engine process fails.
The conventional recovery path uses a cold restart. NVIDIA noted that this can require model weights to be loaded from storage into high-bandwidth memory and kernels to be compiled before the engine resumes operation.
Shadow Engine Recovery addresses the recovery of an engine process rather than a failed GPU, server rack, data center or power system. The supplied source does not claim that it repairs failed hardware or restores an entire facility.
No universal recovery-time benchmark, model size, cluster configuration or hardware platform was included in the supplied description. The stated timing is therefore NVIDIA’s general characterization of recovery occurring “within seconds.”
Google Cloud and Anyscale introduced experimental gVisor sandboxing for distributed Ray clusters, available through Ray 2.58 APIs. #
Google Cloud and Anyscale introduced an experimental Ray library that uses gVisor to provide isolated execution environments inside distributed Ray clusters. Ray is a distributed-computing runtime used to coordinate trainers, inference engines, rollout workers and other components in artificial-intelligence workflows.
Each high-level sandbox is represented as a Ray Actor. Ray’s scheduler chooses the node, reserves central processing unit and memory resources, and manages placement, while the Actor manages the sandbox lifecycle and gVisor supplies the isolated runtime environment.
Beginning with Ray 2.58, framework authors and researchers can manage sandboxes with the same Ray application programming interfaces used for other cluster resources. A sandbox can be created from an Open Container Initiative-compatible image, assigned CPU and memory limits, and accessed through normal Ray Actor calls from anywhere in the cluster.
The sandbox API supports environment creation, networking configuration, environment variables, working directories and command execution. Google positioned the capability for workloads such as dynamic rollouts, code generation and multi-turn tool interaction, while explicitly describing the library as experimental.
- Google Cloud
- Anyscale
- Ray
- Ray 2.58 APIs
OpenAI confirmed data-center executive Chris Malone departed as some whole-facility leasing projects restarted under other leadership. #
An OpenAI spokesperson confirmed that Chris Malone was no longer with the company and said OpenAI had reorganized its infrastructure organization earlier in 2026. Malone reportedly left during the week before the disclosure.
Malone joined OpenAI in March 2025 and oversaw data-center construction associated with Stargate, the infrastructure initiative involving OpenAI, Oracle and SoftBank. The sources describe him as a key executive responsible for data-center construction and expansion.
People familiar with the matter said OpenAI was restarting some internal infrastructure projects by leasing entire data-center facilities. Those projects had been transferred to other leaders, so the restart was not described as remaining under Malone’s control.
The supplied reports do not identify the number, location, power capacity or lease value of the restarted facilities. They also do not establish that the projects were operational at the time of Malone’s departure.
Google Cloud launched Gemini Enterprise for Legal and Gemini Enterprise for Financial Services, adding governed agents, reusable skills, and secure Model Context Protocol connectors. #
Google Cloud introduced separate Gemini Enterprise offerings for legal work and for capital-markets and corporate-banking workflows. Both products combine reusable domain skills, governed agents, secure connections to enterprise systems and partner integrations.
Gemini Enterprise for Legal includes skills for contract review and redlining, playbook creation, regulatory monitoring, legal research and Data Subject Access Request fulfillment. Its Model Context Protocol connectors inherit permissions from document-management systems, case repositories, research services and other legal applications.
Gemini Enterprise for Financial Services includes a Google-managed Financial Research agent with more than 50 foundational skills. Google said the agent provides confidence scores, explicit methodologies, auditable data snapshots and source citations, and can be used through the Gemini Enterprise application or integrated through Agent-to-Agent application programming interfaces.
Financial-services Model Context Protocol connectors preserve existing data entitlements, meaning licensed and permissioned information remains subject to its original access controls. Google also identified systems integrators, legal-technology companies and financial-technology providers as participants in the respective partner ecosystems.
- Google Cloud
- Model Context Protocol
Figure AI launched an index and robot-training dataset, pledging over $1 billion for data and compute during the next year. #
Figure AI announced an index and a robot-training dataset. The supplied source does not provide the index methodology, dataset size, covered tasks, licensing terms or the number of robot-operation examples.
The company committed to invest more than $1 billion in data and compute over the following 12 months. The commitment combines two spending categories and does not break out how much will go to computing infrastructure versus data acquisition or preparation.
Figure AI also reported 44,000 weekly active users contributing data. The source does not specify whether every user contributes at the same frequency or volume, or how contributed information is validated and incorporated into robot training.
The stated amount is a company commitment for future spending, not disclosed expenditure already completed. No accelerator model, cloud provider, data-center location, power capacity or procurement contract was identified.
Waymo unveiled a 5-nanometer autonomous-driving ASIC and will begin Munich testing, targeting public ride-hailing service by the end of 2027. #
Waymo introduced a 5-nanometer application-specific integrated circuit designed for autonomous-driving computation. The chip processes sensor data in real time and is intended for Waymo’s fully autonomous ride-hailing platform.
The Alphabet subsidiary said it would begin testing vehicles in Munich within the following weeks. Initial vehicles will be operated manually to create high-definition maps and verify that Waymo’s autonomous-driving software can handle Munich’s specific road conditions.
During the validation phase, trained autonomous-driving specialists will sit in the driver’s seat. Waymo is targeting the launch of public ride-hailing service in Munich by the end of 2027, making Germany its planned first European Union market and third planned overseas market.
Waymo said it had contacted local, state and federal officials regarding the approvals needed to offer service in Germany. The company also said it plans to invest in local fleet operations and create skilled jobs before launching in Munich, without disclosing fleet size or investment value.
