What goes around comes around. The AI's 700 billion debt will start to be settled this week.

Everybody, it's Monday again. I was working overtime all weekend, the competition for domestic voice-over (dubbing) resources is getting fiercer, everyone is scrambling for projects.

No need to say much about the IPO subscription market, recently it's been all about all-in bets, today's grey market trading for Shenghong and the IPO launch for Xizhi are both very good targets. It's a shame that the HL grey market price for Shenghong was suppressed too hard by a few large orders, the price just couldn't go up. Those who sold at the high opening price were lucky, but tomorrow it might still rise, so it's hard to say whether not selling was good or bad.

On the US stock side, earnings season enters a dense period this week. But more important than the earnings numbers themselves, these numbers need to answer a question ignited by DeepSeek in January last year but still unresolved: Is the $700 billion that tech giants have poured into AI actually making money?

I. The beginning of the story starts with DeepSeek

Many say DeepSeek impacted companies like Microsoft and Meta, but they are not the same type of players at all. DeepSeek is an AI model research company focused on training large models. Microsoft, Meta, Amazon, and Google are spending CapEx not on the models themselves, but on computing infrastructure. They make money from compute rental fees and traffic monetization, essentially an infrastructure business.

At the time, DeepSeek achieved near-top-tier model performance at a lower cost, which ignited a question: If strong models can be trained with less computing power, is the ceiling for AI's demand for computing power not that high?

Most counterarguments at the time focused on "the reduction in training costs will instead lead to an explosion in computing demand at the inference stage". This reasoning is logically sound, but logic holding up is one thing, whether the numbers materialize is another. This fortnight's earnings reports are the first real data test of this debate.

II. Two tiers, two layers of verification

This year, the combined capital expenditure of the four major supercomputing players is close to $700 billion: Google $175-185B, Amazon $200B, Meta up to $135B, Microsoft expected to exceed $120B for the full year, up over 60% from 2025 overall. With such massive investment, the market needs answers on two levels.

1. The reality of AI demand

First, look at Tesla, IBM, Texas Instruments, Lam Research. This essentially verifies whether AI demand is translating into real orders.

Texas Instruments' data can tell us if chip demand is substantially recovering; Lam Research's order situation reflects downstream wafer fabs' real judgment on future capacity; IBM's volume of AI enterprise contracts is the first B-side validation of the logic of exploding inference-side demand. The core of Tesla's earnings this time is not deliveries, but CapEx guidance—how much of its promised $20 billion in capital expenditure is actually being deployed, whether the Robotaxi timeline is delayed, these directly affect the market's pricing of the AI hardware demand narrative.

2. Is the spending paying off?

Next week, focus on Meta, Microsoft, Amazon, Google. Two numbers are key:

First, cloud revenue growth rate. Whether Azure, Google Cloud, AWS are growing fast enough determines if CapEx is strategic investment or a money-burning black hole. Current signals are positive—Google Cloud grew 48% YoY in Q4 2025, Amazon Bedrock's API calls in Q1 were already 3x the full-year 2025 volume. But Microsoft also disclosed $80 billion in Azure orders unable to be fulfilled due to power shortages, demand is being constrained by supply, whether growth rates reflect real demand will only be clear next week.

Second, whether CapEx guidance will be raised again. Morgan Stanley's expectation for hyperscale capital expenditure in 2027 is already about 15% higher than market consensus. If earnings reports raise it again next week, the market must choose: either believe inference demand is really exploding; or think free cash flow erosion is too severe. Good news needs to be super-exceeding expectations to push prices up, bad news just needs to be slightly below expectations to trigger a sell-off, this asymmetry is the biggest risk next week.

The long-term logic for 'picks and shovels' stocks hasn't changed, but with valuations already back to pre-war highs, overall, entry timing is more important than position size.

$Alphabet(GOOGL.US) $NVIDIA(NVDA.US) $Amazon(AMZN.US) $Meta Platforms(META.US) $Tesla(TSLA.US) $IBM(IBM.US) 

The above content represents personal views only and does not constitute investment advice. Give Vivian a follow, wishing everyone that what you buy goes up and what you sell goes down~