Google, Microsoft, Amazon Earnings: AI Spending Strains Cash Flow, Custom Chips Shine

This quarter, all three cloud giants posted record capex: Alphabet's Q4 capex was $44.9B, doubling year-over-year, with full-year guidance raised to $195B-$205B and explicit plans for significant increases in 2027; Microsoft spent $41B in the quarter, up 69.4% year-over-year, totaling $145.3B for the year; Amazon spent $54.2B in the quarter, up from $32.1B a year ago, maintaining full-year guidance of $200B.

The cost is deteriorating free cash flow: Alphabet turned negative for the first time in history at -$5.9B, prompting $49.6B in equity issuance, $20.3B in bonds, and a $40B ATM plan; Amazon's free cash flow turned negative for the first time to -$7.6B, due to a $66.1B year-over-year increase in property and equipment purchases for AI investments; Microsoft's free cash flow was $19.6B, down 23% year-over-year, but remained positive.

All three companies signal that 'demand exceeds supply,' with capex primarily directed to GPU compute and data centers. Amazon's CEO specifically noted that rising memory prices are one driver of the capex boom.

Google's TPU system sales are now officially included in Cloud revenue. This quarter, Google Cloud revenue was $24.8B, up 82% year-over-year, with an operating margin of 35.6% and a backlog of $514B. The CFO stated that TPU systems have begun shipping, with a small portion of revenue recognized in 2026 and the majority in 2027. Google is also considering providing $44B in guarantees for third-party data center leases, backing about ten projects totaling 2.4 gigawatts of capacity to drive TPU sales.

Amazon's Trainium series is accelerating. Trainium 3 server Q3 shipment targets were raised by 20%-30%, with nearly all capacity reserved, and some customers have already reserved Trainium 4. AWS revenue this quarter was $42.2B, up 36.7% year-over-year, the fastest growth in 18 quarters, with an operating margin of 39%, contributing nearly 61% of operating profit, and a backlog of $496B. AI business annualized revenue exceeds $25B, with triple-digit year-over-year growth, plus $225B in committed orders.

Microsoft's custom chip Maia 200 is deployed at scale, and Cobalt 200 is in early preview, but they are primarily used for internal cost reduction, with no significant revenue contribution yet. In contrast, Microsoft faces a disadvantage in cost structure.

Google's Gemini 3.5 Pro has been delayed three times, directly due to programming and complex task capabilities not meeting internal standards. The trigger was Anthropic's Claude Opus 4.7 surpassing Gemini 3.1 Pro across benchmarks like coding and tool use, which Google acknowledged as a 'fundamental difference in engineering capability.' Sergey Brin's internal memo urged a decisive pivot to agent technology. Google has positioned Gemini 4 as a 'next-generation architecture rebuilt around agents' and is prioritizing TPU resources accordingly.

In parallel, Google has launched three mid-tier models—3.6 Flash, 3.5 Flash-Lite, and Cyber—with price cuts, shifting focus to reducing token consumption and inference costs. However, Google is particularly cautious about transforming Gemini into an agent, as about 80% of revenue comes from online ads, fearing agents bypassing search interfaces could hollow out its commercial foundation. On the earnings call, Google responded that 'Gemini is strengthening monetization, not diluting it.'

Microsoft is naturally transitioning to the agent phase based on its office ecosystem. CEO Nadella emphasized 'separating the control layer from the model itself.' Microsoft 365 Copilot has 30 million paid seats, and it launched a fully autonomous agent called 'Autopilot.' Microsoft also introduced its first reasoning model, MAI Thinking-1, with over 11,000 models on its cloud platform, and Foundry customers grew fivefold. Amazon similarly emphasizes a multi-model strategy, with Bedrock serving over 100,000 customers, and new customers in the past six months exceeding the total from the first two years after launch.

All non-core, non-AI businesses are being rapidly divested. Google DeepMind disbanded the AlphaFold author team, shifting research focus entirely to Gemini, and formed an AI coding task force. DeepMind's VP of Research admitted the strategy has shifted from scientific projects to competing with OpenAI and Anthropic.

Amazon AWS scaled back several AI model projects just before earnings. The AGI division saw multiple changes: former head Rohit Prasad left in December 2025, AGI Lab closed after David Luan's departure, and Frontier Model Research, led by Pieter Abbeel, became the core.

Microsoft announced a new strategy at Build: not building the smartest model, but building an operating system for agents, with resources concentrated on cloud and AI.

This quarter's earnings show that overseas tech companies' AI race has reached a stage where profits must be accounted for. Whether selling chips, selling MaaS, cutting organizations to make way for AI, or using financial maneuvers to boost profits, all expansion must be justified by facts.

Custom ASIC chips are becoming the next major battleground for cloud vendors, providing sustainable profit points for massive capex. DIGITIMES Research predicts ASIC server shipments will grow 64.2% in 2026, far exceeding GPU servers' 43.8%.

Microsoft's extension of depreciation to 25 years lowered FY2027 capex guidance to ~$175B, essentially an accounting treatment rather than a real slowdown in investment, improving free cash flow and book profits.