Big Tech's $745B AI Spending Strains Cash Flow

Key Takeaways

• <ul><li>Alphabet, Amazon, Meta, Oracle, and Microsoft plan roughly $745 billion in AI-related capital expenditures this year, straining free cash flow across Big Tech.</li><li>Broadcom's AI chip revenue grew 221% year over year, and its CEO expects AI revenue to hit $230 billion by fiscal 2028.</li><li>Nvidia reported Q2 fiscal 2027 revenue of $96 billion and projects 70% revenue growth in fiscal 2028.</li><li>Michael Burry is shorting Micron, betting that the AI memory chip boom will follow a cyclical pattern and reverse.</li><li>Meta launched Muse Spark Open, an 816B mixture-of-experts coding model trained on NVIDIA GB300 NVL72, delivering up to 50K tokens per second per GPU.</li><li>OPPO partnered with Manus to bring agentic AI to its Find X10 Series smartphones via AI Mind Pilot's Agent Mode.</li><li>A Fortanix survey found 41% of enterprises blocked an AI deployment over security concerns, though 61% still send sensitive data to public cloud AI services.</li><li>U.S. Bancorp says third-party AI agents are not yet moving customer deposits between banks due to logistical complexity.</li><li>Custom XPUs are emerging as a way for hyperscalers to lower upfront costs and manage AI data center spending.</li><li>Certified Adaptive Selection (CAS) can identify the top-k algorithms with fewer evaluations, certifying the top three after only 630 of 3,150 queries in one test.</li></ul>

Capital shortage is the biggest AI bottleneck say Broadcom and Intel CEOs

Hyperscalers like Alphabet, Amazon, Meta, Oracle, and Microsoft face negative free cash flow from massive AI spending. Alphabet reported negative free cash flow for the first time, and Amazon cited a $66.1 billion increase in capital expenditures. CEOs of Broadcom and Intel say capital constraints are the biggest bottleneck for AI data center build-outs. Companies are turning to custom chips, or XPUs, to lower upfront costs and keep spending in check.

Capital limits AI data center growth while custom chips offer a path forward

Building AI data centers requires huge capital, and long lead times are holding back companies from building as fast as they want. Alphabet, Amazon, Meta Platforms, Oracle, and Microsoft all face similar cash flow challenges from rising AI spending. Custom chips called XPUs offer lower upfront costs and better price-performance on specific tasks. Broadcom leads the XPU market, and companies like Marvell Technology could also see rapid growth from XPU sales.

Big Tech's $745 billion AI spending fuels suppliers Nvidia Broadcom Arista

Amazon, Alphabet, Microsoft, and Meta are set to spend roughly $745 billion on capital expenditures this year for AI. Nvidia, Broadcom, and Arista Networks supply the GPUs, custom silicon, and networking gear that turn that spending into working AI capacity. Nvidia reported Q2 fiscal 2027 revenue of $96 billion and expects 70% revenue growth in fiscal 2028. Broadcom's AI chip revenue grew 221% year over year and its CEO expects AI revenue to reach $230 billion by fiscal 2028.

Michael Burry shorts Micron over AI memory cycle and valuation risks

Famous investor Michael Burry is betting against Micron, a major memory chip maker. He is concerned that the current boom in high bandwidth memory for AI processors will follow a historical cyclical pattern and reverse. Burry also worries about circular financing deals supporting AI data center build-outs and the rise of Chinese memory-chip makers. Micron is trading at a low price-to-earnings ratio despite strong growth, but Burry sees risks ahead.

Michael Burry bets against Micron on memory chip cycle and financing risks

Michael Burry, known for his successful bet against the housing market, is now shorting Micron stock. He believes the memory chip market's current surge in demand for high bandwidth memory used in AI processors will eventually reverse. Burry also raises concerns about circular financing deals in the AI industry and Chinese memory-chip makers potentially ending the global memory shortage sooner. Despite Micron trading at low earnings multiples, Burry sees significant risk factors for the company.

OPPO teams with Manus to add AI agents to smartphones

OPPO partnered with AI platform Manus to bring agentic AI to its next-generation smartphones. The new features will appear in AI Mind Pilot's Agent Mode on the OPPO Find X10 Series with ColorOS 17. Users can ask the AI to handle multi-step tasks like planning a trip or creating a presentation from photos and files. The system keeps local data on the device and asks for user permission before sharing content externally.

Qwen 3.8 27B training sped up 6x on AI Runtime

A blog post explains how to fine-tune the Qwen 3.8 27B model on AI Runtime six times faster. Even with GPU utilization at 98 percent, training time dropped from 7 hours 44 minutes to 1 hour 17 minutes. Key techniques include installing the correct kernels, improving the data pipeline, and reducing excess padding in batches. These changes cut the number of batches by about 36 percent and required relatively little effort.

Restaurant owner turns to ChatGPT to name a bun

Nina Zhang, co-owner of Noodles by Nina in Denver, used ChatGPT to find a more appealing English name for her egg custard bun. She also uses the AI chatbot for menu adaptation, trademark checks, inventory, and handling workplace conflicts. Many local restaurateurs now use AI tools to help run their businesses amid thin profit margins. Some in the industry see this as a creative threat, while others call it a helpful free assistant.

Survey finds 41% of enterprises blocked AI over security

A Fortanix survey of 300 AI leaders found that 41 percent of enterprises blocked an AI deployment due to security concerns. Although 93 percent worry about exposing proprietary data to AI providers, 61 percent still send sensitive data to public cloud AI services. Sixty-four percent have deployed closed-weight frontier AI models on-premises, and 97 percent would adopt leading models if securely available on-premises. Confidential Computing was the top factor that would increase confidence in AI deployment.

Meta launches Muse Spark Open coding model on NVIDIA GB300

Meta launched Muse Spark Open, a coding-focused 816B mixture-of-experts model with 42B active parameters and a 1M token context window. Trained on NVIDIA GB300 NVL72 Blackwell Ultra, it supports agentic coding workflows with tool calling and deep codebase understanding. The model processes text, images, audio, and video through a 1.8B perception encoder in a single pass. Running on NVIDIA GB300 NVL72, it delivers up to 50K tokens per second per GPU and up to 600 tokens per second per user.

U.S. Bancorp says AI agents are not moving consumer deposits yet

U.S. Bancorp reported that third-party AI agents are not yet moving customer deposits between banks. Credit card balances rose 11.5% year over year, and the net charge-off rate fell to 3.29%. Consumer deposits reached a quarterly record, though deposit costs increased seven basis points. The company's CEO said the logistics of AI agents automatically moving cash are very complicated, so no near-term impact is expected.

New method selects top algorithms with fewer evaluations

Researchers introduced Certified Adaptive Selection (CAS), a method that identifies the exact top-k set of algorithms with fewer evaluations. CAS keeps guaranteed bounds on Expected Shortfall and stops testing when remaining uncertainty cannot change the result. The prediction-guided version CAS-PG certified the exact top three after 630 of 3150 queries on one dataset. Without historical data, a transcript-only variant required 1.28 to 1.37 times the minimum number of evaluations.

Framework evaluates credit assignment in modular financial reinforcement learning

A new evaluation framework addresses credit assignment challenges in modular financial reinforcement learning. The framework combines equal-information reconstruction controls, fixed-utility comparisons, and prespecified downstream guardrails. Tests found that residual-based allocation is reproduced by an ordinary loss-budget control, while direct shortfall supervision improved prediction but failed outcome preservation. The approach identifies three evaluation failures and offers a reusable methodology for systems operating under partial observability and economic risk constraints.

Sources

• Not Memory. Not Energy. This Is the Biggest Bottleneck in Artificial Intelligence, and These Are the Companies That Can Benefit.
• Not Memory. Not Energy. This Is the Biggest Bottleneck in Artificial Intelligence, and These Are the Companies That Can Benefit.
• Big Tech Will Spend $745 Billion on AI This Year. These Stocks Get Paid on Every Dollar.
• Michael Burry Is Short This AI Darling. Should You Be Worried?
• Michael Burry Is Short This AI Darling. Should You Be Worried?
• OPPO partners with Manus to integrate artificial intelligence agents into smartphones
• Reducing Qwen training time by 6x on AI Runtime
• A Chinese restaurant owner wanted a better name for a bun. So she turned to ChatGPT.
• Fortanix Survey Finds 41% of Enterprises Blocked an AI Deployment over Security
• Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
• U.S. Bancorp Says AI Agents Aren’t Moving Consumer Deposits Yet
• Certify the Decision, Not Every Score: Exact Top-k Selection from Few Evaluations
• Equal-information evaluation of credit assignment in modular financial reinforcement learning