AI Compute Cloud & Neoclouds

Amazon.com is reportedly considering moving roughly $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle financed by outside investors and then leasing the hardware back, a structure that would make part of its AI buildout more asset-light as capital spending is expected to reach $220 billion this year. AWS revenue rose 37% to $42.2 billion in the second quarter, its fastest growth in more than four years, and contract backlog reached $496 billion, while CEO Andy Jassy said the company still lacks enough capacity to meet demand. Trailing-12-month free cash flow swung to negative $7.6 billion from positive $18.2 billion a year earlier as infrastructure spending accelerated. Amazon trades at 23.64 times forward earnings, and Tigress Financial recently raised its target to $385 from $315, arguing earnings from the investment cycle are approaching an inflection point where they grow faster than operating capital. The proposal also raises the question of residual chip value, since Grace Blackwell will eventually be superseded by Vera Rubin while Amazon assumes semiconductor generations remain useful for at least five years.

Matt Garman, CEO of Amazon.com Inc.'s AWS, said concerns over an AI spending bubble are overstated, telling a16z's podcast with Raghu Raghuram on Thursday that Amazon does not rely on one customer for most AWS capacity and that "we feel really good about the spend we're making now." Garman said demand is concentrated in production workloads such as computing, storage and AI inference, and that a diversified approach means not all capacity is bundled up in one customer, comparing the market to venture capital where failed startups can be offset by successful investments. His comments come as AWS highlights efforts to make AI more affordable, and as Amazon is reportedly in talks to transfer about $8 billion worth of Nvidia Corp.'s Grace Blackwell AI chips to an investor-backed special-purpose vehicle that could raise funds through debt issuance while Amazon leases the chips back. Amazon CEO Andy Jassy raised the company's 2026 capital expenditure forecast by $20 billion to $220 billion in July, citing rising memory chip prices and strong demand for AI and AWS infrastructure. Separately, billionaire investor Ray Dalio warned that rising AI-related borrowing, higher interest rates and pressure to turn paper wealth into cash could bring the AI bubble closer to bursting, calling the boom a "classic bubble" and comparing the current AI rally to the late 1920s.

Australian AI data center operator Firmus, which is backed by Nvidia, has withdrawn its planned $5B initial public offering, citing market volatility and conditions. The company said its board determined the terms of the proposed offering did not adequately reflect the strength of its business and long-term growth outlook, and concluded that proceeding was not in the best interests of the company and its shareholders. Firmus had planned to raise $5B in the IPO, pricing shares at A$11 apiece, which would have made it the second-largest new share sale in Australia's history and valued the company at around $30.6B. The company will now pursue capital from the private markets and consider alternative public and private market options. Firmus unveiled a $2B funding round backed by Nvidia, Coatue Management, Blackstone, and Jane Street in August, bringing its total equity raised over the preceding year to more than $3B and its valuation to over $10.5B. Investors began pulling their orders on Wednesday as they lost confidence in the deal, after being told on Tuesday about escrow arrangements that would have allowed more than half the stock to be sold by existing investors from day one.

Tianyang Technology announced before market open today via its official WeChat account that it has signed a GPU Procurement Agreement with Company X to purchase 20,000 NVIDIA RTX PRO 5500 Blackwell units. Boosted by the news, the company's closing price that day was 16.16 yuan per share, up 12.38% from the previous trading day. A Cailian Press reporter interviewed multiple industry insiders and learned that the NVIDIA RTX PRO 5500 Blackwell has not yet officially gone on sale, and its official pricing has not been determined. The domestic quote for this GPU may exceed 100,000 yuan per unit. Based on a purchase volume of 20,000 units, the aforementioned agreement could be worth more than 2 billion yuan, but practitioners bluntly said the order has barely begun to take shape, and getting the goods is estimated to happen no earlier than the first quarter of next year. Xia Hailong, a lawyer at Shanghai Shenlun Law Firm, said a preliminary estimate puts the order amount at potentially several billion yuan. If it reaches the financial threshold requiring disclosure, it should be disclosed through statutory channels in accordance with the law, and publishing it only through a WeChat official account is clearly non-compliant. A representative from Tianyang Technology's securities department responded that specific details are inconvenient to disclose, and if the relevant disclosure threshold is met, there will be subsequent disclosure. Previously, Tianyang Technology had already disclosed large-scale computing power projects, including a project cooperation agreement signed in May with Qiming Xinghan and others for an overall computing power procurement scale in the range of 3.5 billion to 4 billion yuan, a joint venture established in June with Capital Online with a capital contribution of 70 million yuan for a 70% stake, and a Computing Power Cloud Service Agreement signed with Yunli Zhihui with a total tax-inclusive amount of 303 million yuan, which was raised to 402 million yuan in July. The semi-annual report shows that as of the end of June this year, the company's fixed assets were approximately 275 million yuan. The two major contracts with Qiming Xinghan and Yunli Zhihui recorded zero sales revenue recognized in the current period and cumulatively. In the first half of the year, non-GAAP net profit was negative 6.9296 million yuan, turning to a loss year-on-year. Net cash flow from operating activities was approximately negative 341 million yuan. Monetary funds were approximately 508 million yuan, and short-term borrowings were approximately 615 million yuan.

Amazon.com Inc. plans to deploy two million additional NVIDIA GPUs during 2027 and 2028 as part of its AI infrastructure buildout, alongside a $1.9 billion investment in its Delivery Service Partner program in 2027 that is mostly directed toward driver wages. That driver-pay commitment, which Amazon says will lift driver pay to a national average of nearly $24 an hour, brings total funding for the eight-year-old program to $21.7 billion. The company also completed a £4.25 billion sterling bond offering on September 14, 2026, across four maturities of 3, 6, 12, and 19 years, broadening its financing sources as AI spending raises capital requirements. In the second quarter, Amazon posted $200.6 billion in total revenues, up 20% from Q2 FY25, with AWS revenue jumping 37% year-over-year to $42.2 billion, while 2026 capital outlays are projected at close to $220 billion and free cash flow has slipped into negative territory. Separately, Amazon is raising base starting wages for qualifying full-time employees in core U.S. operations by $1 to $20 an hour effective September 27, a move it says reflects a commitment of over $1.5 billion.

Nvidia-backed AI infrastructure startup Firmus Technologies is expected to delay its planned initial public offering in Australia, according to a Bloomberg report citing people familiar with the matter. Firmus is instead in talks with existing investors and others for a private funding round, though order-taking for the IPO closed as scheduled on Thursday morning with no clear indication of the price or deal structure. Deliberations are ongoing and details could still change, the people said, and a spokesperson for Firmus did not immediately respond to a Bloomberg request for comment. Bloomberg reported last month that Firmus was in talks to raise about $10 billion in financing ahead of its planned Australian IPO. Firmus began as a Bitcoin miner in Tasmania in 2019 and now operates seven AI factories across Australia, Singapore, Indonesia, and Malaysia.

Microsoft has merged its three reporting segments into two, Agents and Infra and Devices and Consumer, starting with fiscal 2027, a change that puts its AI investment case in sharper focus. Restated figures show the Agents and Infra segment, which groups Azure with Microsoft 365 cloud and industry solutions, generated $268.1 billion in revenues and $136.4 billion in operating income in fiscal 2026, an operating margin above 50% even at the peak of the buildout, and management expects first-quarter fiscal 2027 Agents and Infra revenues of $75.15-$75.75 billion. The spending is enormous: fiscal fourth-quarter capital expenditures were $41 billion, roughly two-thirds of it on short-lived assets mainly CPUs and GPUs, first-quarter capex is expected to exceed $50 billion, and a lease reclassification has lifted the calendar 2026 estimate to about $175 billion, while free cash flow fell to $19.6 billion from about $25.6 billion a year earlier and Microsoft Cloud gross margin fell year over year to 65%. The strongest offset is visible future demand, with commercial remaining performance obligation up 84% to $678 billion, about 30% of it expected to be recognized as revenue within 12 months, Azure revenues exceeding $100 billion for the first time in fiscal 2026, and first-quarter Azure growth projected at 44-45% in constant currency against about 17% for Microsoft 365 commercial cloud, supported by more than 30 million paid Copilot seats and new usage-based billing. For fiscal 2027, Microsoft expects double-digit growth in revenues and operating income, full-year operating margins slipping by less than a point and free cash flow staying positive, and because so much of the spending goes to short-lived hardware the company can slow purchases quickly if demand weakens, though rising depreciation and margin pressure remain real risks. Amazon raised its 2026 cash capex outlook to about $220 billion, with AWS revenues up 37% to $42.2 billion in the second quarter and a backlog of $496 billion, while Alphabet lifted its 2026 capex guidance to $195-$205 billion as Google Cloud revenues surged 82% to $24.8 billion and its cloud backlog reached $514 billion, with slightly more than half expected to convert into revenue within 24 months, though Alphabet's free cash flow turned negative in the second quarter.

Microsoft has joined major tech peers in issuing large new debt packages to fund AI infrastructure buildouts, borrowing to finance data centers and AI hardware rather than relying solely on existing cash flows and reserves. Economists and credit analysts are flagging the rapid expansion of AI related debt as a possible source of bubble like conditions. The company, which carries a US$3.9 trillion market cap, is tying itself more tightly to the AI data center cycle, a signal that management is confident enough in AI demand to lock in financing for chips, power and campuses. Analysts already frame Microsoft around huge AI capex of roughly US$175b to US$190b in 2026, with the risk that this infrastructure only works if AI and cloud usage stay strong enough to support revenue growth and margins under heavier capital intensity. Investors will watch how future results link reported AI revenue, Azure growth and gross margin trends to higher interest expense and capex, particularly when Microsoft rolls out its new FY27 reporting around Agents and Infra.

Nvidia is set to benefit from a planned $40 billion SpaceX fundraising led by Apollo aimed at securing exclusive access to Nvidia chips. SpaceX is seeking US$40b in new capital to more than double its Nvidia-powered compute footprint for large scale AI and space data projects, effectively treating Nvidia as the default compute layer for those efforts. Separately, OpenAI has chosen Nvidia hardware over Cerebras to run its new Ultrafast AI model, reinforcing Nvidia's role in large model training and its full stack platform anchored by CUDA and TensorRT. The key test for investors is whether these headlines turn into disclosed, multi year capacity commitments that show up in data center segment demand and long term supply agreements, with concrete contract details from SpaceX and further large scale model deployments from OpenAI that explicitly reference new Nvidia platforms such as Blackwell or Rubin. The developments sit inside a wider build out of AI plumbing across data centers and cloud platforms.

Two widely accepted narratives about global artificial intelligence are unraveling as market data show U.S. models are often cheaper per completed task than Chinese competitors, while enterprise AI spending is stabilizing despite surging usage. Martin Chorzempa, a senior fellow at the Peterson Institute, wrote in a social media post Tuesday that benchmarking data from Artificial Analysis indicates top U.S. models from developers like OpenAI, Anthropic, and Google tend to be more efficient with tokens when accomplishing equivalent tasks, even though Chinese models may offer lower upfront pricing per token. Chorzempa noted that industry observers routinely focus on technical papers from Chinese labs highlighting architectural improvements while overlooking what U.S. firms have behind the scenes that would cut down on their number one cost, alongside access to more efficient hardware chips. He also pointed out that while open-weight models can be downloaded for free, most commercial enterprises rely on cloud infrastructure, where Chinese labs are reportedly asking for a 30% cut from cloud providers, passing additional expenses on to enterprise users. Separately, Chorzempa cited analysis from Ara Kharazian, lead economist at AI finance platform Ramp, showing that corporate AI spend actually fell 5.2% in recent tracking even as token consumption volumes reached all-time highs, with businesses becoming increasingly adept at picking the right cost and capabilities tradeoffs and leveraging fierce price competition between major U.S. labs.