- Blackwell Ultra (B300) is shipping: inference costs have fallen ~40% compared to B100
- Project DIGITS delivers personal 1-petaflop AI compute at $3,000 in a desktop form factor
- NIM microservices are in general availability across AWS, Azure, GCP, and on-premises
- NVIDIA Cosmos physical AI platform has 300+ robotics and automotive partners in 2026
- Inference cost reduction will continue at roughly 50% per year through the decade
NVIDIA's 2025 GPU Technology Conference set the trajectory for enterprise AI infrastructure through the rest of the decade. Now that Blackwell hardware is shipping in volume and the software ecosystem has matured, the strategic implications are considerably clearer than they appeared on stage in San Jose.
Blackwell Ultra in production: what the cost curve means
The B300 series began reaching hyperscalers in Q3 2025. Compared to H100s from two years prior, Blackwell Ultra delivers roughly 4× the inference throughput per dollar. That compounding improvement in price-performance is what drives API pricing down across every major provider: OpenAI, Anthropic, Google, and their enterprise competitors all route inference through NVIDIA silicon.
The practical implication for 2026 budget planning: any projection built on 2024 cost assumptions is already outdated. AI workflows that required human augmentation purely for cost reasons are worth re-evaluating. The economics are shifting faster than most annual planning cycles account for.
Project DIGITS: local AI hits a real price point
Project DIGITS generated more consumer excitement than any NVIDIA product since the original RTX launch. The GB10 Grace Blackwell Superchip combines a Grace CPU with a Blackwell GPU in a Mac Mini-sized enclosure, delivering enough compute to run Llama 3.1 405B in quantised form locally.
The $3,000 price point matters for organisations with data sovereignty requirements. Regulated industries: healthcare, legal, financial services: now have a credible on-premises inference option that does not require a rack of A100s or a data centre contract. Two DIGITS units linked via NVLink handle 405B models in full precision.
NIM microservices: production deployment simplified
NVIDIA Inference Microservices reached general availability in late 2025 and represent the company's most important software move of the year. NIM packages popular foundation models: Llama, Mistral, Stable Diffusion, and NVIDIA's own models: as containers with pre-built API compatibility, GPU optimisation, and driver management handled automatically.
For engineering teams, the value is measurable: deploying a new model in a NIM-compatible environment takes minutes rather than days. The consistency of the REST interface across model versions and hardware generations also reduces the testing burden when switching or upgrading models.
Physical AI and the longer horizon
GTC 2025's most forward-looking thread was NVIDIA's push into physical AI: systems that perceive and act in the real world via the Isaac robotics platform and Cosmos simulation environment. In 2026 this is beginning to materialise in production: warehouse robots, autonomous vehicles, and industrial inspection systems are running on NVIDIA-trained AI at significant scale.
For software teams, this remains a horizon watch rather than an immediate procurement question. But the infrastructure choices made now: which cloud, which AI platform, which inference runtime: are increasingly shaped by NVIDIA's architectural decisions made at events like GTC.
Frequently asked questions
What was the biggest announcement at NVIDIA GTC 2025?+
The Blackwell Ultra GPU (B300 series) was the headline hardware announcement, promising roughly 1.5× the inference throughput of the original B200 at a lower per-token cost. Project DIGITS: a personal AI desktop delivering one petaflop of local compute at $3,000: attracted nearly as much attention.
What is Project DIGITS and is it available in 2026?+
Project DIGITS is a compact desktop computer built around the GB10 Grace Blackwell Superchip. Announced at GTC 2025 at $3,000, it reached general availability in mid-2025. It can run 70B-parameter models in full precision and 200B models in quantised form. Two units can be NVLink-connected for larger workloads.
What are NVIDIA NIM microservices?+
NVIDIA Inference Microservices (NIM) are containerised, pre-optimised packages for deploying AI models in production. They abstract away GPU tuning and driver management, letting enterprise teams deploy foundation models with a single API call on any NVIDIA hardware generation.
How do GTC 2025's announcements affect AI tool pricing in 2026?+
Directly. The Blackwell Ultra's cost-per-token improvements have translated into API price cuts across every major AI provider in 2025–2026. Workflows that were too expensive to run at scale in 2024 are now economically viable. Budget planning for AI in 2026 should assume continued cost reduction at roughly 50% per year.
What is NVIDIA Cosmos and who is it for?+
NVIDIA Cosmos is a physical AI development platform for building AI systems that operate in the real world: warehouse robots, autonomous vehicles, and industrial inspection. It combines the Omniverse simulation environment, pre-trained world foundation models, and the Isaac robotics framework. It is relevant for organisations building embodied AI products rather than software-only applications.