NVIDIA Partners with Amazon to Expand from Accelerators to Processors, Networking, Open Source Models, and Robotics
NVIDIA CEO Jensen Huang stated in a declaration about expanding cooperation with Amazon Web Services that demand is outpacing all forecasts, and the collaboration is extending from accelerators to processors, networking, open source models, and robotics. CFO Colette Kress provided a framework for a revenue increase of about 70% for FY2028 during the earnings call, noting that customer forecasts indicate a doubling, but the company is guiding only half of what can be supplied. Revenue for the quarter was $96.2 billion, with data center revenue at $89 billion. The company has never written a guidance of 'earning $1 trillion next year'; a 70% growth rate corresponds to a scale of about $690 billion to $700 billion.
Vera Rubin has entered the mass production and distribution window, with the demand-supply gap first quantified: demand is about 140%, while supply is locked at only 70%. Bloomberg reported that some of the largest customers have been notified that many configurations of servers equipped with Vera Rubin and Grace Blackwell will see price increases of over 15% for shipments in early 2027, primarily due to high bandwidth memory. Gross margin guidance dropped from 75% for the quarter to about 74% for the third quarter, with a bottom of about 71% to 72% in Q4 FY2027, and returning to 72% to 73% in FY2028. Kress stated that memory price increases are higher than previously assumed and will be even greater, with price increases fully reflected in the next fiscal year.
Amazon Web Services announced plans to deploy an additional 2 million NVIDIA accelerator cards from 2027 to 2028, covering Blackwell Ultra, Rubin, and Rubin Ultra, adding to the previous plan of 1 million cards; among these, 100,000 cards will be used for high-security computing facilities for the U.S. government. Kress stated that the deployment extends from the current quarter to Q2 FY2029. The bottlenecks identified include packaging, storage, silicon photonics, substations, land, and electrical work, not customer willingness. Huang emphasized that buying cards is akin to building computing facilities, not just purchasing a computer.
NVIDIA's stake in SpaceX was previously disclosed in a 13F filing with an approximate market value of $21 billion. Musk stated that ground and in-orbit computing is shifting to the exclusive Vera Rubin architecture. The statement that 'just SpaceX alone has generated over $30 billion in profit' is presented as realized profit, without corresponding confirmation from the company. There exists a customer financing flywheel: the company has substantial equity, cloud commitments, and residual value arrangements to support labs with insufficient balance sheets. The complete wording regarding employment and 'AI tax' did not appear in the line-by-line correspondence of this earnings report and AWS statement, and cannot be considered a key point this time.
The buyers are procurement officers looking to lock in cabinets for cloud factories and sovereign projects from 2027 to 2028; the sellers are those whose memory and packaging have already been booked into next year's supply chain. Funding is shifting from trading concerns about bubbles back to supply quota premiums. Beneficiaries are those who can enhance the value of each gigawatt system through whole machine sales; those under pressure are the gross margins that must first absorb memory price increases before being passed on. The events were driven by the earnings report and AWS's expanded orders, not by trillion-dollar profit promises.
Source: Public Information
ABAB AI Insight
Huang wrote the gap as a business model. Customers want 200%, but he only commits to 70%, with the difference being the right to raise prices. The 15% is not a romantic markup on chip catalog prices, but rather HBM rising first, followed by cabinets. Gross margins dropped from 75% to a corridor of 72%, acknowledging that part of the excessive profits already belongs to storage manufacturers. Locking in three-year memory contracts shifts volatility from the current profit statement to the settlement date in 2028.
The capital path is that cloud factories nearly fill operational cash flow with capital expenditures, and NVIDIA uses equity, guarantees, and private infrastructure funds to turn card buyers into financed parties. AWS's 2 million cards are turning the pre-sale of 2028 factories into a deployment timeline. The stake in SpaceX turns customers into assets on the balance sheet, and the exclusive architecture ties the next round of in-orbit computing back to the same NVLink. The money does not come from token retail, but from gigawatt-level prepayments and residual value commitments.
Benchmarks include TSMC's advanced packaging quotas, SK Hynix selling out its HBM capacity for 2026, and Cisco writing routers into internet GDP. The current phase is quota management during expansion: those who can squeeze out more CoWoS and HBM will write 70% as a higher number. Control lies in storage and packaging, not in analysts' trillion-dollar slogans.
Structurally, pricing power is shifting. Pricing power is moving from GPU gross margins to memory BOM and shipping years. The mechanism is: computing factories buy by gigawatt, while memory is contested by wafers; memory must first become the major capital expenditure before chip manufacturers can write 15% into next year's cabinets. Treating 'earning $1 trillion next year' as a direct quote is writing supply options as already booked profits.
ABAB News · Cognitive Law
- The gap itself is pricing power.
- Customers want to double, but guidance only writes half of what can be supplied.
- Memory must rise first for chips to write 15% into next year's cabinets.