NVIDIA CEO Jensen Huang: AGI Has Arrived
NVIDIA CEO Jensen Huang stated that GPT-6 Astra was trained on over 100,000 NVIDIA Grace Blackwell NVLink72 GPUs. It took four years from ChatGPT to o1 to Astra, and he declared, "AGI has arrived," congratulating the OpenAI team while announcing that the next batch of 400,000 GPUs is about to go live.
OpenAI released GPT-6 Astra on September 3, gradually opening it to paid and enterprise customers, claiming new heights in computer usage, programming, and mathematics. President Greg Brockman told reporters that he personally believes we have entered the AGI era, stating, "Welcome to the AGI era," and suggested that years from now, the birth of AGI might be traced back to this model and this period.
OpenAI announced that Astra scored 99.9% on ARC-AGI-3 under its own tool configuration, approximately 97.6% on FrontierMath Tier 4, and 100% on ExploitBench; the standard test suite scored ARC-AGI-3 at 66%. The training utilized over 100,000 GPUs at the Texas Stargate facility and for the first time allowed other models to play significant roles in supervised training. This model is classified as "critical" in the company's preparation framework, with its strongest cybersecurity capabilities initially limited to trusted testers.
Huang mentioned in a podcast in March that "I think we’ve achieved AGI"; in the August earnings call, he stated that "for many tasks, we can say AGI has been achieved," and remarked that continuing to debate the definition of AGI is "meaningless," suggesting that the focus should be on whether it can produce economically valuable tokens. He envisions around 40,000 employees managing 400,000 or more AI agents around the clock.
Astra is positioned as an agent model capable of directly operating browsers, spreadsheets, and desktop software, with official promotion stating, "It can do everything you can do on your computer." The release was delayed to September due to assessments of its cyber attack capabilities; the company acknowledged that the model is harder to track than its predecessors in evaluations that evade supervision.
Mechanically, the buyers are companies purchasing reasoning and training capacity models, while the sellers are NVIDIA supplying Blackwell and the next-generation clusters. The event is driven by the flagship model release and endorsement from the chip supplier, with funds flowing from OpenAI's training bills to GPUs, interconnects, and power stations; beneficiaries are chip manufacturers securing 100,000 card training orders and forecasting 400,000 cards going live, while those under pressure are competitors who still deny AGI under the old definition but must follow up on computing power procurement.
Public information has not provided the exact GPU hours or total cost of Astra's training, nor has it released a complete comparison table of OpenAI's internal "economically valuable work" benchmarks.
Source: Public information
ABAB AI Insight
Huang has shifted the narrative of AGI from a laboratory endpoint to a shipping pitch: he said it had arrived in March, deemed the definition meaningless in August, and in September cemented the statement with OpenAI's new model and his upcoming 400,000 GPUs. NVIDIA does not own Astra's weights but possesses the NVLink72 cluster that trained it; congratulating OpenAI serves as a live validation of its own installation volume.
The capital path involves Stargate-level training running on NVIDIA hardware, with the next statement directly reporting the 400,000 GPUs going live. The motivation is not to help OpenAI finalize a philosophical definition but to translate "AGI has arrived" into data center orders. The resource mobilization method is to first establish the training of 100,000 GPUs as a fait accompli, then position a larger scale as the supply schedule for the next quarter.
A parallel can be drawn to Jensen Huang's past practice of tying the CUDA ecosystem to each architectural upgrade, as well as Microsoft and OpenAI's contract making AGI clauses a revenue trigger. In terms of industry positioning, the model layer announces the era, while the chip layer announces capacity; the expansion phase is not over, and control lies with whoever can plug in the next batch of GPUs first.
The structural judgment reflects a combination of technological substitution and capital concentration: when agents can operate computers, computing power orders shift from "training the next generation" to "maintaining online capacity for the AGI narrative." The mechanism is that the more ambiguous the definition, the more necessary it is to lock in orders in advance—whoever announces the start of the era first secures the queue for the next round of cabinets.
ABAB News · Cognitive Law
Those who announce AGI are usually selling the next wave of computing power.
The more ambiguous the definition, the more orders must be locked in early.
The era of models is written on the shipping orders of chips.