

OpenAI Cuts Frontier Model Pricing as Inference Commodifies
OpenAI released three GPT-5.6 tiers on July 9: Sol, Terra, Luna priced at $5, $2.50, $1 per million tokens, undercutting Anthropic and signaling pricing war.
Chips, data centers, and the trillion-dollar buildout
Compute, not capital, has become the binding constraint of the AI industry. Hyperscalers have announced over $660B in data center capex, frontier labs sign multi-year GPU leases worth tens of billions, and a single training cluster can consume the power of a small city. This hub tracks the silicon roadmaps, data center buildouts, supply chain moves, and compute economics that decide which AI companies can actually scale.


OpenAI released three GPT-5.6 tiers on July 9: Sol, Terra, Luna priced at $5, $2.50, $1 per million tokens, undercutting Anthropic and signaling pricing war.


NVIDIA's Rubin ships across hyperscalers in H2 2026, ending GPU scarcity but intensifying power grid constraints now favoring only three frontier labs.


Data Center Power Coalition launches July 1 with 12 partners to standardize on-site power for AI data centers facing 5-year grid transformer backlogs.


OpenAI's Jalapeño chip offers 50% cost savings on inference, undermining Nvidia's GPU dominance as frontier labs shift to custom silicon.


OpenAI and Broadcom ship custom LLM inference chip with 50% lower costs versus NVIDIA, targeting production deployment Q4 2026 with gigawatt-scale rollouts.


50,000 attendees, Kawasaki's 8-DOF humanoid, and ABB's physical AI standard reveal enterprises moving from evaluation to procurement in manufacturing.


xAI Grok 4.3 achieves sub-1% hallucination while costing 2-5x less than GPT-4o Reasoning, now available on Amazon Bedrock with configurable reasoning effort.


SandboxAQ secures $500 million CHIPS R&D award to discover AI-driven materials replacing toxic forever chemicals in semiconductor fabs


Hydra Host's $100M Series A, backed by Nvidia and ARK Invest, builds a GPU marketplace across 50+ data centers to solve AI compute fragmentation.


Anthropic's Fable 5 export control shutdown reveals the fatal flaw in cloud-first AI: when governments act, tenants lose and infrastructure owners survive.


ASUS GB300 ExpertCenter Pro runs 1-trillion-parameter models on a desktop, bringing 20 PFLOPS to enterprise without renting cloud infrastructure.


Qualcomm is in talks to buy AI chip startup Tenstorrent at up to $10 billion, a move that would reshape who controls the future of AI inference hardware.


Lawrence Livermore scientists warn AI power use is unsustainable in Science, as US data centers hit 42GW and ionic chips emerge as a 100x efficiency path.


FERC is set to rule by June 30 on who funds grid expansion for AI data centers, a decision that could raise power bills for 65 million Americans.


AMD's strategy chief outlined four AI buildout walls at SuperAI: copper deficits, turbine backlogs, power grid delays, and an emerging memory shortage.


AI data center demand doubled to 42 GW since 2023 as power grid capacity, not chip supply, now binds global AI infrastructure growth.


TrendForce Q1 2026 data shows TSMC commands 72.3% of global foundry revenue at $35.9B, while Samsung trails at 6.5%, the widest gap ever recorded.


Meta's 2 billion dollar Rivos acquisition is not delivering on custom AI chip goals, undercutting its plan to break free from Nvidia.


Oracle's Q4 cloud revenue hit $9.9B with OCI up 93%, but a $70B capex plan for FY2027 drove shares down 7%, signaling AI infrastructure's true cost.


Nvidia's Vera CPU becomes orderable by Chinese cloud providers for August delivery, targeting $20B in FY2027 revenue as China market re-entry begins.


Big Tech is building private power plants for AI data centers, bypassing utilities as electricity demand from AI grows past 224 terawatt-hours a year.


Google forced Crusoe off Project Jade, a 1.8GW Wyoming data center, after raising concerns about cost and construction timetable, Bloomberg reports.


MIT spinout Ferveret brings nuclear cooling to AI chips, delivering 35 percent more tokens from the same power with zero water use.


Nvidia's 800VDC transition may slip past 2028 as suppliers report no clear roadmap for the next power architecture shift in AI data centers.


NEMA, ASHRAE, and PNNL launch a unified AI data center standard backing 800VDC power as AI facilities now demand up to 80MW of power each.


Broadcom, Apollo, and Blackstone's $35B AI XPV Platform targets 20 gigawatts of compute through 2028, backed by frontier labs Anthropic and OpenAI.


Nvidia's Rubin racks will draw 230 kW each, and a new Digitimes analysis confirms electricity, not GPUs, is now what limits AI expansion.


Google will pay SpaceX $920 million a month for 110,000 Nvidia GPUs through June 2029 to feed Gemini Enterprise demand it failed to forecast.


Anthropic will pay SpaceX 1.25 billion dollars a month for over 220,000 GPUs through May 2029, a 45 billion compute bill that dwarfs its own revenue.