Top Tech News Today, September 4, 2026: Amazon, Google, Microsoft, Nvidia, OpenAI, Tesla & More

It’s Friday, September 4, 2026, and the AI boom stopped looking theoretical overnight. OpenAI released GPT-6 Astra and declared the arrival of the AGI era. Tesla put wheel-free Cybercabs on Austin streets and almost immediately drew a federal safety probe. Meanwhile, Crusoe vaulted to a $30 billion valuation, Congress moved to inventory rogue AI agents, and a Saudi lab released a 428-billion-parameter Arabic model built on Chinese weights.

AI is no longer just competing for your screen. It is moving into chips, data centers, cars, robots, wearables, and national security. Today’s biggest technology stories capture that shift in real time: OpenAI is pushing frontier models into critical-level cyber capabilities, Broadcom is gaining ground in custom AI chips, Tesla is testing the regulatory limits of autonomous transportation, and billions of dollars are pouring into the infrastructure required to train the next generation of intelligent machines.

At the same time, governments are scrambling to catch up. Europe is wrestling with its dependence on U.S. cloud giants, American lawmakers are proposing unprecedented limits on advanced AI, and China is aggressively policing AI-generated content. The message from today’s headlines is hard to miss: the AI race is no longer simply about who builds the smartest model. It is becoming a global contest over compute, energy, hardware, regulation, security, and who ultimately controls the machines.

From frontier AI models and billion-dollar startups to autonomous vehicles, cyber risk, and governments racing to regulate the technology, here are the tech stories that actually moved the stack today.

OpenAI Launches GPT-6 Astra as Its Most Powerful AI Model Yet, With Critical-Level Cyber Capabilities

OpenAI has unveiled GPT-6 Astra, its newest frontier AI model and the company’s most ambitious attempt yet to push artificial intelligence beyond chat and into complex, end-to-end professional work. OpenAI says Astra sets new internal benchmarks across software engineering, computer use, web browsing, science, cybersecurity, and other professional tasks. The model is initially rolling out to a limited group of organizations before expanding to ChatGPT Plus, Pro, Business, and Enterprise customers, as well as developers through the OpenAI API, Microsoft Azure, and Amazon Bedrock. API pricing starts at $10 per million input tokens and $50 per million output tokens.

The bigger development may be Astra’s security profile. OpenAI says the model is its first broadly deployed system to reach the “Critical” cybersecurity capability threshold under the company’s Preparedness Framework. According to OpenAI, Astra can, under some conditions, discover previously unknown vulnerabilities and develop methods for exploiting hardened systems without continuous human direction. That capability forced OpenAI to introduce tighter internal isolation, encrypted checkpoints, expanded trajectory monitoring, and additional alignment reviews. The milestone shows how quickly frontier models are shifting from tools that assist humans to systems that can independently complete sophisticated technical workflows. It also raises the stakes for AI safety as the same capabilities that help defenders identify vulnerabilities could potentially be misused by attackers.

Why It Matters: GPT-6 Astra pushes frontier AI deeper into autonomous professional work while simultaneously demonstrating why cybersecurity and AI safety are becoming inseparable from the race for more capable models.

Source: TechStartups via OpenAI.

Microsoft AI Ships MAI-Transcribe-2 at $0.10 an Hour, Undercutting OpenAI and Google

Microsoft AI released MAI-Transcribe-2, a speech-recognition model it says is faster, more accurate, and cheaper than rival services from OpenAI, Google, and ElevenLabs. The introductory price is $0.10 per audio hour through the end of 2026, down about 72 percent from the $0.36 an hour Microsoft charged for the first model in the line five months ago. The system covers 60 languages, up from 43 in June’s 1.5 release, and adds diarization, configurable styles, and word-level timestamps. Microsoft says it ranks first on the FLEURS multilingual benchmark with a 5.2 percent average word-error rate and, on Artificial Analysis measurements, runs 10 times faster than GPT-Transcribe, seven times faster than ElevenLabs Scribe v2, and five times faster than Gemini 3.5 Transcribe.

Speech-to-text is one of the few AI markets where you can compare price, latency, and language coverage on a spreadsheet. Microsoft is using that transparency to push its in-house MAI models through Foundry, MAI Playground, and OpenRouter rather than relying only on OpenAI models inside Azure. For call-center platforms, media companies, and meeting-note startups, a drop from $36,000 to $10,000 on 100,000 hours of audio is large enough to force a rebid. It also shows how quickly first-party Big Tech models are turning yesterday’s premium AI feature into a utility, leaving startups to compete on workflow, compliance, and domain data rather than raw transcription quality.

Why It Matters: Microsoft just turned speech recognition into a 10-cent commodity, squeezing AI audio startups and OpenAI’s own transcription stack at once.

Source: VentureBeat.

U.S. Opens Tesla Cybercab Safety Probe Hours After Autonomous Vehicle Launch

Tesla’s push toward fully autonomous transportation is facing an immediate regulatory test. The U.S. National Highway Traffic Safety Administration has opened a probe into nearly 1,000 Tesla Cybercabs, focusing on whether the vehicles comply with federal motor vehicle safety standards. The investigation began shortly after Tesla started putting Cybercabs on public roads in Austin, Texas. Unlike conventional cars, the purpose-built autonomous vehicle has no steering wheel, brake pedal, accelerator pedal, or traditional mirrors.

Tesla told regulators it had self-certified the Cybercab as compliant with applicable federal standards, a normal process for automakers. But the absence of manual controls creates an unusual regulatory situation because many existing vehicle standards were written around cars operated by human drivers. Federal transportation officials have proposed modernizing some of those rules for autonomous vehicles, but Tesla’s commercial deployment is moving faster than the regulatory framework. The probe could therefore become an important test case for the broader robotaxi industry, including companies such as Waymo, Zoox, and other autonomous vehicle developers. If regulators determine that purpose-built autonomous vehicles need exemptions or new certification rules before large-scale deployment, the consequences could extend well beyond Tesla.

Why It Matters: The Cybercab investigation could help determine how quickly fully driverless vehicles without traditional controls can move from experimental fleets into mass-market transportation.

Source: Reuters.

Adobe Names Anil Chakravarthy CEO as Generative AI Reshapes the Creative Software Market

Adobe has named longtime executive Anil Chakravarthy its next president and CEO, placing new leadership at the top of one of the technology company’s most directly exposed to the generative AI revolution. Chakravarthy will take over on December 1, while longtime CEO Shantanu Narayen will transition to executive chair. Chakravarthy joined Adobe in 2020 after leading Informatica and, most recently, Adobe’s Customer Experience Orchestration business and worldwide field operations.

The leadership transition comes at a pivotal moment for Adobe. Photoshop, Illustrator, Premiere, and the company’s other creative products remain industry standards, but AI-native tools are increasingly generating images, video, advertising copy, and design assets without requiring traditional editing workflows. Adobe has responded by embedding its Firefly generative AI models across its product portfolio and expanding AI-powered marketing products such as GenStudio. Chakravarthy now inherits the challenge of protecting Adobe’s enormous installed base while convincing customers that AI strengthens rather than replaces its creative software ecosystem. Adobe also faces pressure to prove that the transition can produce durable revenue rather than simply higher computing costs. Narayen helped turn Adobe from a packaged-software company into a subscription giant. Chakravarthy’s task will be navigating an equally significant transition from conventional SaaS to AI-native software.

Why It Matters: Adobe’s next CEO will be judged largely on whether the company can remain the default creative platform as generative AI changes how digital content is produced.

Source: The Wall Street Journal.

AI Infrastructure Startup Crusoe Raises More Than $3 Billion at a $30 Billion Valuation

Crusoe has reportedly raised more than $3 billion in new funding at a valuation of roughly $30 billion, another enormous capital injection into the infrastructure underpinning the AI boom. The round comes only about 10 months after Crusoe raised $1.38 billion at a roughly $10 billion valuation, meaning the company’s valuation has approximately tripled in less than a year. Atreides Management and Valor Equity Partners are co-leading the financing, with Mubadala Capital among the participants.

Crusoe’s transformation illustrates how quickly the economics of computing have changed. The company started in 2018 using stranded and flared natural gas to power cryptocurrency mining operations. It later pivoted toward AI infrastructure and has become one of the emerging “neocloud” providers supplying GPUs, data centers, and specialized computing capacity to major technology companies. Customers reportedly include Meta, Microsoft, and OpenAI. Crusoe has also signed a five-year cloud agreement worth roughly $13 billion to provide Jane Street with GPU infrastructure. The funding reflects investors’ growing belief that the largest opportunities created by AI may not belong solely to model developers. Companies supplying electricity, data centers, networking, cooling, chips, and compute capacity are becoming critical pieces of the AI economy.

Why It Matters: Crusoe’s $30 billion valuation shows how AI infrastructure providers are becoming major technology companies in their own right as demand for compute keeps rising.

Source: Reuters.

EU Defense Officials Push Back Against Plans to Reduce Reliance on Amazon, Microsoft, and Google Cloud

European defense officials are resisting parts of a proposed EU technology framework designed to reduce dependence on American cloud and AI providers. The European Commission’s planned Cloud and AI Development Act seeks to strengthen Europe’s digital sovereignty by encouraging sensitive public-sector workloads to move toward European-controlled technology infrastructure. But defense officials in several countries warn that replacing services provided by Amazon, Microsoft, Google, and other U.S. companies too quickly could create operational and security problems.

The dispute exposes a difficult trade-off facing Europe. Policymakers increasingly view cloud infrastructure and AI computing capacity as strategic assets, particularly as governments rely more heavily on software platforms operated by foreign companies. Yet European alternatives remain smaller and, in some areas, technologically behind U.S. hyperscalers. Defense officials are particularly concerned about interoperability with NATO systems and military hardware that already depends on American cloud services. The proposed rules would classify government workloads by sensitivity, with the strictest restrictions expected to apply to only a small percentage of services. Still, the debate illustrates how cloud computing is becoming part of geopolitical strategy. Europe increasingly wants technological independence, but moving away from U.S. infrastructure could carry high costs and security complications.

Why It Matters: The debate shows that cloud and AI infrastructure are increasingly being treated as strategic national-security assets rather than ordinary enterprise technology.

Source: Financial Times.

U.S. Lawmakers Propose Permanent Ban on Artificial Superintelligence and Pause on Advanced AI

Sen. Bernie Sanders and Rep. Greg Casar have announced legislation that would permanently prohibit the development and deployment of artificial superintelligence while temporarily pausing certain advanced AI development until federal safety regulations are established. The proposed Ban Artificial Superintelligence Act would also direct the U.S. government to pursue international agreements to prevent superintelligent systems from being developed elsewhere. Sanders is among the highest-profile U.S. politicians so far to call for an outright prohibition rather than simply stronger regulation of frontier AI.

The proposal comes as AI laboratories release increasingly autonomous systems that can code, browse, operate computers, conduct scientific research, and execute multi-step tasks. OpenAI’s GPT-6 Astra release has sharpened the debate because the model reached the company’s “Critical” cybersecurity capability level. Supporters of stronger restrictions argue that development is outpacing governments’ ability to establish oversight. Critics counter that broad restrictions could push research into other countries and weaken U.S. competitiveness. The legislation faces substantial political hurdles, but its significance extends beyond its likelihood of passage. Washington’s AI debate is moving beyond questions about copyright, misinformation, and employment toward the possibility that some capabilities might be considered too dangerous to develop at all.

Why It Matters: The proposal marks a major escalation in the U.S. AI policy debate, shifting the conversation from regulating powerful models to potentially banning an entire category of future AI systems.

Source: Axios.

Microsoft Unveils Project Zenith for On-Device AI Development on 64GB PCs

Microsoft named its developer-optimized Windows configuration Project Zenith and tied it to a new class of PCs with at least 64GB of unified memory and 250GB/s of memory bandwidth. The first devices ship with AMD Ryzen AI Halo silicon; AMD showed a miniature Zenith PC on stage at IFA. Microsoft says the machines will come with a preconfigured, distraction-free Windows setup and a curated toolchain so developers can run 30-billion-parameter-plus models locally and without metered cloud tokens. File Explorer will surface extensions, hidden files, and full paths by default. Windows Terminal and Visual Studio Code are pinned to the taskbar, with GitHub Copilot, PowerToys, WinAppCLI, and Windows Dev Skills preinstalled.

The move aims to pull high-end AI experimentation off the GPU cloud and onto a Windows workstation. Local 30B-class models matter for startups that cannot absorb token bills, for teams handling sensitive code, and for vendors trying to prove that Windows can still be the default environment for AI-native development as Apple and Linux-based stacks chase the same buyers. Microsoft said additional Zenith devices on other silicon will follow in the coming months, turning the label into a hardware-plus-software program rather than a one-off SKU. If OEMs price these machines within reach of independent developers, the practical effect is a new on-device inference tier sitting between phones and the data center.

Why It Matters: Microsoft is turning high-memory PCs into a first-class platform for local frontier-scale models, cutting developers’ dependence on paid cloud inference.

Source: Windows Central.

Europe’s Tech Chief Says U.S. AI Guardrails Are Becoming Inevitable Despite Regulatory Divide

European Commission Executive Vice President Henna Virkkunen says the apparent divide between European and American AI policy may be smaller than it appears. Speaking around this week’s G20 innovation summit in North Carolina, Virkkunen argued that the United States is gradually creating many of the same safeguards Europe favors, even though Washington continues to reject the EU’s comprehensive regulatory approach. Rather than passing a single framework similar to the EU AI Act, the U.S. is building oversight through state laws, court rulings, executive actions, voluntary testing frameworks, and national-security restrictions.

The distinction matters because AI governance is increasingly becoming a competitive issue between major economic blocs. U.S. officials and technology executives have warned that overly broad rules could slow innovation while Chinese AI companies continue advancing. Europe argues that pre-deployment testing, transparency requirements, and accountability mechanisms are necessary as AI systems become more autonomous. Yet both sides increasingly agree that frontier models with powerful cyber, biological, or autonomous capabilities may require additional safeguards. The emerging debate is therefore shifting from whether AI should be regulated at all to which risks require oversight and who should enforce it. Global companies may ultimately face overlapping regulatory expectations, even if governments don’t formally copy Europe’s AI Act.

Why It Matters: Despite political disagreements over regulation, the U.S. and Europe appear to be converging on the idea that increasingly powerful AI models will require stronger safeguards.

Source: Axios.

Broadcom’s AI Chip Business Surges as Custom Silicon Challenges Nvidia’s Grip on AI Compute

Broadcom delivered another signal that the AI semiconductor boom is expanding well beyond Nvidia’s GPUs. The chip and infrastructure software giant reported fiscal third-quarter revenue of $29.6 billion, up 86% from a year earlier, while quarterly profit more than tripled to roughly $13.1 billion. A major driver was demand for custom AI accelerators from large technology companies building their own computing infrastructure. Broadcom’s AI semiconductor revenue reached $16.7 billion during the quarter, according to The Wall Street Journal, with custom chips accounting for roughly 73% of that total.

The numbers highlight a structural change underway inside the AI infrastructure market. Hyperscalers including Google, Meta, OpenAI, and other large AI companies increasingly want chips optimized around their own workloads rather than relying exclusively on general-purpose accelerators. Broadcom has emerged as one of the biggest beneficiaries of that shift by helping technology companies design specialized silicon and networking systems. The company expects AI semiconductor revenue to continue climbing aggressively as customers expand data centers and next-generation training clusters. Nvidia remains the dominant supplier of AI compute, but custom silicon could gradually reduce the industry’s dependence on any single chip architecture. The bottleneck is increasingly becoming manufacturing capacity, advanced packaging, networking, and memory rather than simply access to GPUs.

Why It Matters: The AI chip market is evolving from a largely Nvidia-centered ecosystem into a broader race involving custom accelerators built specifically for the largest AI platforms.

Source: The Wall Street Journal.

Saudi HUMAIN Unveils 428-Billion-Parameter Arabic AI Model Built With China’s MiniMax

PIF-backed Saudi AI company HUMAIN launched humain-m3, a 428-billion-parameter mixture-of-experts model commissioned from Chinese lab MiniMax and further pre-trained on more than one trillion tokens of Arabic-native text. HUMAIN said the model posted the highest average score among frontier systems tested on seven public Arabic benchmarks covering understanding, knowledge, language quality, factuality, and retrieval-augmented generation. A research preview is live on HUMAIN Node. Open weights are planned under the MiniMax Community License after additional safety training, targeted for next month. MiniMax shares in Hong Kong jumped after the announcement.

The release is as much a geopolitical product as a language model. Gulf sovereign AI strategies have leaned heavily on U.S. cloud and chip partners; building a flagship Arabic model on a Chinese open base tests how far U.S. allies will go to assemble sovereign stacks when Washington restricts advanced chips and models. CEO Tareq Amin framed the project as closing a representation gap for hundreds of millions of Arabic speakers. For startups in the region, Node access and forthcoming weights could lower the cost of Arabic agents, government systems, and media tools. For U.S. and European vendors, it is another sign that the open-weight supply chain now runs through Shanghai and Riyadh as well as San Francisco.

Why It Matters: A Saudi sovereign AI champion just put a Chinese open model at the center of the Arabic-language frontier, widening the map of who trains national AI stacks.

Source: ZAWYA.

Qualcomm Backs Indian Smart-Ring Startup Ultrahuman in $70 Million Funding Round

Indian wearable technology startup Ultrahuman has raised $70 million in new financing as it attempts to turn the smart ring into a broader computing platform. Qualcomm Ventures joined the round alongside Labcorp, Alpha Wave, Blume Ventures, Nexus Venture Partners, and Alteria Capital. The financing includes roughly $65 million in equity and $5 million in debt and values the Bengaluru-based company at approximately $365 million, according to TechCrunch.

Ultrahuman has built its reputation around health and fitness tracking, competing with products such as Oura Ring and Samsung’s Galaxy Ring. Its next phase is more ambitious. The company wants future rings to run software directly on the device, potentially supporting AI interactions, health applications, games, and other services without relying entirely on smartphones. Qualcomm’s involvement is particularly notable because the chipmaker has extensive experience building low-power processors for mobile and wearable devices. Smart rings remain a relatively small category compared with smartwatches, but their discreet form factor and continuous biometric monitoring make them attractive as potential interfaces for AI assistants. If computing power, battery efficiency, and sensors continue improving, rings could become part of a broader shift toward ambient computing that reduces dependence on screens.

Why It Matters: Ultrahuman’s funding suggests investors and chipmakers increasingly see AI wearables as potential computing platforms, not just health trackers.

Source: TechCrunch.

Trezor Says ShipMonk Data Breach Exposed Another 67,000 Hardware-Wallet Customers

Crypto hardware-wallet maker Trezor has significantly expanded the number of customers believed to have been affected by a breach involving logistics provider ShipMonk. Trezor said approximately 67,000 additional U.S. customers had their personal information exposed, bringing the total to nearly 14,000 customers previously identified. The newly discovered records date from 2019 through 2021 and reportedly include names, email addresses, telephone numbers, shipping addresses, and order numbers.

There is no indication that cryptocurrency private keys or wallet assets were directly compromised, but the exposed information creates a different kind of threat. Hardware-wallet customers are particularly attractive targets for phishing, impersonation, and physical-extortion attempts because attackers may infer that people who purchased a device hold digital assets. The breach also illustrates the expanding security perimeter facing technology companies. Trezor can secure its own infrastructure and cryptographic hardware while still being exposed through fulfillment companies, customer-support vendors, and other third parties. Supply-chain breaches have become an increasingly common way for attackers to obtain information from organizations whose core systems may otherwise be well protected. For crypto companies, where customers can permanently lose assets after a successful phishing attack, even seemingly ordinary shipping data can carry unusually high security consequences.

Why It Matters: The incident shows how third-party logistics providers can become weak links in otherwise security-focused technology ecosystems, particularly when leaked customer data can enable highly targeted attacks.

Source: The Block.

China Removes 5.6 Million Pieces of Content in Sweeping Crackdown on AI Misuse

Chinese authorities say they have removed more than 5.61 million pieces of unlawful or rule-violating online content during a nationwide campaign targeting misuse of artificial intelligence. Regulators also took action against more than 49,000 accounts and dealt with over 2,400 websites and applications. The Cyberspace Administration of China said the campaign focuses on AI-generated misinformation, impersonation, violent or vulgar material, content affecting minors, and other abuses involving generative AI tools.

China has moved faster than many governments in building an extensive regulatory structure around generative AI. Providers already face obligations involving personal data, content controls, and security, while rules that took effect in 2025 require AI-generated text, images, audio, and video to contain visible or embedded identifiers. The scale of the latest enforcement campaign demonstrates that Beijing is increasingly moving from writing rules to actively policing AI-generated content. China is simultaneously encouraging aggressive domestic AI development, creating an unusual policy combination: rapid technological expansion alongside extensive centralized oversight of how models are deployed and what they produce. The approach contrasts sharply with the lighter regulatory strategy promoted by the United States and could become an alternative model for governments concerned about deepfakes, impersonation, misinformation, and other AI-related harms.

Why It Matters: China’s enforcement campaign provides an early look at what large-scale government policing of AI-generated content may look like as generative media spreads globally.

Source: Xinhua.

India’s ISRO Successfully Launches EOS-05 Earth-Observation Satellite After Seven-Month Mission Hiatus

India’s space program returned to flight Friday with the successful launch of the EOS-05 Earth-observation satellite aboard a GSLV-F17 rocket from the Satish Dhawan Space Centre in Sriharikota. The launch occurred at 2:55 a.m. local time, with ISRO reporting that the vehicle successfully injected the satellite into its intended orbit. EOS-05 is designed to provide high-resolution observation capabilities for applications including agriculture, disaster management, weather monitoring, environmental analysis, and national-security operations.

The mission is important beyond the satellite itself. India is expanding its position in the global space economy while simultaneously opening more of its domestic launch and satellite industry to private companies. Reliable government launch capability remains a foundation for that strategy because commercial startups depend heavily on national infrastructure, regulatory support, and access to launch services. Earth observation is also becoming one of the most commercially significant segments of space technology as satellite imagery feeds agriculture platforms, insurance models, climate monitoring, defense intelligence, logistics, and AI-based geospatial analysis. India already operates one of the world’s largest fleets of remote-sensing satellites, and higher-frequency imaging could strengthen both domestic applications and commercial services.

Why It Matters: The EOS-05 mission reinforces India’s growing role in Earth observation and comes as the country works to build a larger commercial space and satellite ecosystem around ISRO’s capabilities.

Source: India Today.

Quantum Motion Expands Series C to Scale Silicon-Based Fault-Tolerant Quantum Computers

London-based Quantum Motion has announced a second close of its Series C financing as the startup works to build fault-tolerant quantum computers using conventional silicon semiconductor manufacturing techniques. New investors include Imec.ventures, Lansdowne Partners, Sony Innovation Fund, S3 Ventures, and Inkef, joining a financing previously led by DCVC and Kembara. The capital will support custom silicon development, semiconductor manufacturing partnerships, and international expansion, including a new U.S. laboratory in Maryland.

Quantum Motion is pursuing one of the most consequential questions in quantum computing: whether quantum processors can eventually be manufactured using the same CMOS fabrication infrastructure that produces conventional computer chips. Many existing quantum architectures depend on highly specialized components that make systems extraordinarily expensive and difficult to manufacture at scale. Quantum Motion argues that its silicon spin-qubit approach could eventually cut hardware costs by orders of magnitude while leveraging decades of semiconductor manufacturing expertise. The company is also participating in Stage B of DARPA’s Quantum Benchmarking Initiative, a U.S. effort examining whether emerging quantum architectures can reach economically useful performance. Quantum computing remains years away from widespread commercial deployment, but manufacturability is becoming just as important as raw qubit performance.

Why It Matters: If quantum processors can be built using established semiconductor manufacturing infrastructure, scaling fault-tolerant quantum computers could become dramatically more practical and affordable.

Source: The Quantum Insider.

IFA 2026 Shows AI Moving Beyond Chatbots Into Robots, Smart Devices, and ‘Embodied Intelligence’

The IFA technology show opening in Berlin this week is highlighting a major shift in the consumer electronics industry: artificial intelligence is moving out of software interfaces and into physical devices. Showcased products include humanoid robots, AI-powered smart binoculars, augmented-reality diving equipment, next-generation household robots, and other devices that can sense and interact with the physical world. The emerging category is often described as embodied AI or physical AI, where intelligence is paired with cameras, microphones, sensors, actuators, and robotic systems.

That shift could become one of the largest technology transitions following generative AI. Chatbots demonstrated that foundation models can reason across enormous amounts of information, but those systems have largely remained confined to screens. Robotics companies are now attempting to give similar models the ability to perceive environments and take actions in the real world. Falling sensor costs, better vision models, faster processors, and improved robotics hardware are helping accelerate development. Consumer applications may initially appear in specialized products such as home robots, wearables, and smart appliances before expanding into logistics, factories, elder care, construction, and other industries. The challenge will be making these systems reliable enough to operate safely outside controlled environments.

Why It Matters: The next phase of AI may be defined less by smarter chatbots and more by intelligent machines that can see, move, and act in the physical world.

Source: Forbes.

Figure Signs $3.5 Billion AI Compute Deal With Nscale for Up to 100,000 Nvidia GPUs

Humanoid robotics company Figure has signed a massive computing partnership with AI infrastructure provider Nscale that could eventually deploy up to 100,000 Nvidia GPUs. The agreement begins with a $3.5 billion compute commitment and could scale beyond $6 billion. Initial infrastructure based on Nvidia’s Vera Rubin platform is targeted for deployment in the second half of 2027 at a facility in Barstow, Texas. Nscale will also become a shareholder in Figure and serve as the robotics company’s preferred computing provider.

The size of the commitment demonstrates how computationally intensive the race to build general-purpose humanoid robots is becoming. Figure is developing its Helix AI models to allow robots to perceive environments, understand instructions, manipulate objects, and learn increasingly general physical tasks. Training those systems requires enormous amounts of video and sensor data, along with large-scale GPU clusters. Figure recently announced an initiative called Index that it says generates roughly 35 minutes of humanoid training data every second. Physical AI therefore appears poised to create another major source of demand for AI infrastructure alongside language models, image generators, and autonomous software agents. If humanoid robotics succeeds commercially, the compute requirements could become comparable to those of frontier AI labs.

Why It Matters: Figure’s multibillion-dollar compute commitment suggests humanoid robotics is entering the same infrastructure-intensive scaling race that transformed large language models.

Source: Nscale and Figure.