The Night Shift: Overnight Tech News with Jennifer Hudsen • 11:30 p.m. edition • Cutoff 11:40 p.m. EDT / 03:40 UTC
ORIGINAL EDITORIAL CONCEPT — NOT A DATA-CENTER PHOTOGRAPH. Jennifer newsroom art created for The IT Guys visualizes a worldwide flow of AI accelerators into cloud infrastructure. It is not an AWS or NVIDIA image, a real facility, or proof that any announced hardware has been installed.
AWS and NVIDIA Plan 2 Million More AI GPUs—Deployment Is Scheduled for 2027–2028
Status: AWS and NVIDIA jointly announced plans to deploy two million additional NVIDIA GPUs across AWS infrastructure during 2027 and 2028. The language is forward-looking: this is an announced expansion, not two million processors already online. [1]
What happened
The companies say the additional fleet will include Blackwell Ultra, Rubin and Rubin Ultra GPUs distributed across AWS’s global infrastructure, including what they call AI factories. They also plan broader work on Vera CPUs, Spectrum networking, open models, data processing and robotics. Those product and capacity statements come from the companies’ joint release and should be read as their plan, not independent measurement of completed capacity. [1]
This new figure is additional to the more than one million NVIDIA GPUs AWS announced in March would begin entering its regions in 2026. NVIDIA’s release says demand exceeded those earlier expectations. TechCrunch described the new commitment as effectively tripling the previously announced count, while correctly preserving the 2027–2028 deployment window. [1] [2]
No distinct strict-core candidate survived the 9:00:44–11:40:10 p.m. EDT window after duplicate screening. TechCrunch published this follow-on at 4:47 p.m. Pacific / 7:47 p.m. Eastern, three hours and 53 minutes before cutoff, so this issue uses the permitted disclosed near-window fallback. Bloomberg’s 5:07 p.m. Eastern report independently described Amazon adding two million NVIDIA chips over the next two years, but the central deployment plan remains a company announcement. [2] [3]
Why it matters
Two million additional GPUs is a supply-chain and capacity signal, not just a large headline number. Cloud AI services depend on accelerators, CPUs, networking, power, cooling, buildings and software arriving together. A GPU allocation without energized data-center space is not usable cloud capacity, and the announcement does not provide a region-by-region delivery calendar, customer reservation schedule or total contract value. [1]
The mix also spans multiple NVIDIA generations. Blackwell Ultra is closer to current deployment, while Rubin and Rubin Ultra represent later platforms. That means the two-million total should not be treated as two million identical devices arriving at once. The release does not disclose the split among architectures, the number of complete systems or racks, or how many units will be available through ordinary EC2 instances versus dedicated AI-factory arrangements.
The announcement is also strategically revealing. AWS continues to invest in its own Trainium and Graviton silicon, yet it is simultaneously expanding its NVIDIA capacity. For customers, that suggests the near-term cloud market will remain heterogeneous: custom accelerators may improve price-performance for supported workloads, while NVIDIA’s software and hardware ecosystem remains a major demand path. This is analysis from the coexistence of those product lines, not a company promise that either architecture will be cheaper for every workload.
NVIDIA’s same-day quarterly release reported $96.2 billion in revenue for the quarter ended July 26, including $89.0 billion from Data Center. Those are NVIDIA’s unaudited reported results, not revenue from this AWS expansion, and the release does not assign a dollar value to the two-million-GPU plan. [4]
Who is affected
AWS customers: AI labs, enterprises, startups and public-sector buyers could eventually see more accelerator availability. Nothing in the announcement guarantees lower prices, a specific launch date in a particular region or immediate relief for a constrained instance type.
Infrastructure teams: buyers planning 2027–2028 deployments should track which architecture, region, instance family, networking fabric and reservation model actually becomes available. “Two million GPUs” is not a procurement specification.
Power and data-center communities: hardware at this scale implies large supporting investments, but the announcement does not identify every site, grid connection, water requirement or community agreement. Local effects must be evaluated from project-specific records rather than inferred from a global fleet number.
Chip and systems suppliers: the plan signals demand not only for accelerators but also for memory, CPUs, switches, optical links, racks, cooling and electrical equipment. It does not prove that every supplier has secured orders or that every planned unit will arrive on schedule.
What readers should do
- Keep “plan” and “additional” attached. The two million GPUs are scheduled for 2027–2028 and sit on top of the earlier more-than-one-million announcement.
- Do not multiply by a retail GPU price. Cloud deployments are negotiated systems with networking, CPUs, memory, facilities, software and service economics; the companies did not disclose a contract value.
- Wait for product-level availability. Architecture names and global totals do not tell you which EC2 instance, region, quota or reservation model your workload can use.
- Benchmark your own workload. Compare price, availability, software portability, performance and operational overhead across NVIDIA instances and AWS custom silicon when the actual offerings appear.
- For local projects, demand local records. Utility filings, permits, site plans and community agreements are more informative than a worldwide capacity announcement.
Upside, cautions and unknowns
Upside: a successfully delivered fleet could widen access to newer AI infrastructure and reduce some capacity bottlenecks for customers that need NVIDIA’s ecosystem.
Caution: the central source is a joint corporate announcement. The deployment stretches across two future years and multiple product generations, leaving execution, energy, construction and supply-chain risk between the announcement and usable service.
Unknown: the companies did not disclose the purchase value, architecture split, exact regional allocation, rack count, power requirement, customer pricing or how much capacity is already contractually reserved. [1]
Evidence table
| Claim | Evidence state | Boundary |
|---|---|---|
| Two million additional GPUs | Joint AWS–NVIDIA announcement [1] | Planned deployment, not current installed capacity. |
| 2027–2028 timing | Joint release and TechCrunch [1] [2] | No region-by-region delivery schedule disclosed. |
| Blackwell Ultra, Rubin and Rubin Ultra | Named in the joint release [1] | No public split by architecture or complete-system count. |
| NVIDIA quarterly revenue | NVIDIA reports $96.2B total and $89.0B Data Center [4] | Not the value of the AWS expansion. |
Frequently asked questions
Are two million GPUs already running in AWS?
No. The companies say they plan to deploy the additional units in 2027 and 2028. [1]
Is this exactly triple the earlier AWS plan?
The new two million is additional to the more than one million announced in March. “Tripled” is a useful shorthand for scale, but the earlier figure was more than one million, so the precise combined total is greater than three million rather than exactly three million.
Will AWS prices fall?
The sources do not promise a price cut. Capacity, utilization, power, financing, product mix and customer demand all influence cloud pricing.
Does this mean AWS is abandoning Trainium?
No such change was announced. AWS continues to offer custom silicon while expanding NVIDIA capacity; customers should evaluate the actual service options rather than treating the platforms as mutually exclusive.
What would prove progress?
Specific EC2 instance announcements, regional availability, capacity-reservation details, completed facilities, energized clusters and customer use would move the evidence from a broad plan toward deployed service.
Sources
- NVIDIA Newsroom — “AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI”, August 26, 2026; joint primary announcement checked at cutoff.
- TechCrunch — “Amazon just tripled its order of Nvidia chips over ‘surging demand’”, published August 26, 2026 at 4:47 p.m. PDT / 7:47 p.m. EDT; checked at cutoff.
- Bloomberg — “Amazon to Buy 2 Million Nvidia Chips for Data Center Build-Out”, published August 26, 2026 at 5:07 p.m. EDT; metadata and summary checked at cutoff.
- NVIDIA Newsroom — “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027”, August 26, 2026; primary financial release checked at cutoff.
Reporting cutoff: August 26, 2026 at 11:40:10 p.m. EDT / August 27 at 03:40:10 UTC. No distinct strict-core candidate survived the 9:00:44–11:40:10 p.m. window after duplicate screening, so TechCrunch’s 7:47 p.m. report is the disclosed near-window qualifier at three hours and 53 minutes before cutoff. This one subject was checked against every WordPress status, prior 36-hour coverage, earlier Night Shift output, the shared ledger and Jennifer’s knowledge base.