Fears of an AI bubble are resurfacing, and the role of humans as the gatekeepers and auditors for AI is growing precarious.
01
Industry
AI earnings season sees sell-offs, infrastructure expansion faces widespread skepticism
The US tech earnings season in late October 2026 presented a landscape where rapid business growth coexists with drastic capital market sell-offs.
Alphabet's Q3 earnings report released on October 28 showed that Google Cloud revenue surged 103% year-over-year, but free cash flow for the quarter registered a negative $9.1 billion, while the core search advertising business underperformed market expectations. The next day, Alphabet's stock price plummeted 12.6%, and the doubling of cloud business growth failed to offset investors' concerns about runaway cash consumption.
One day later, Amazon delivered a Q3 report card with a highly similar structure: AWS cloud service revenue rose 39% year-over-year to $46 billion (annualized revenue run rate reaching $184 billion), and the group's overall operating profit hit a new all-time high of $28.1 billion. However, in the past twelve months ending September 30, despite operating cash flow growing 29% to $168.3 billion, the company's free cash flow reversed from a net inflow of $14.8 billion in the same period of the previous year to a net outflow of $42.4 billion.
Among these cash gaps, Alphabet's negative $9.1 billion reflects a single-quarter cash shortfall, while Amazon's negative $42.4 billion covers the cumulative outflow over the past twelve months. During this period, Amazon's property and equipment purchase expenditures after deducting asset disposals surged by $94.8 billion year-over-year, and the annual cash capital expenditure is expected to reach as high as $220 billion.
Microsoft's fiscal 2027 Q1 earnings report released on October 28 became one of the few positive cases. Its Azure revenue grew 48% (exceeding its own guidance of 44% to 45%), total group revenue reached $92.3 billion, and it maintained positive free cash flow even with a single-quarter capital expenditure as high as $45.2 billion. The key lies in its huge enterprise software and office subscription base that provides an extremely thick current cash buffer. As a result, its stock price rose 3% sharply after the earnings release.
This shows that the market is currently greatly concerned about large companies' investments in AI infrastructure that exceed their profitable cash flow. They do not deny that there is a current shortage of computing power and that computing centers are a profitable business. This is not a false boom. But their confidence in the future is far less solid.
For these large companies, they must invest heavily in future infrastructure and computing power to maintain competitiveness. But it is hard to say whether these massive investments can eventually be converted into effective returns.
First, at the current growth rate, they may not make much profit. According to Gartner's calculation based on current AI revenue growth, by 2030, the infrastructure investment of AI companies and cloud businesses may reach $7 trillion, total revenue may be $8.5 trillion, and the enterprise and consumer AI revenue attributable to AI data centers will not exceed $3.6 trillion. Considering its operating costs, depreciation, and taxes, its effective profit margin is likely to be lower than 12%. In other words, the profit efficiency is relatively low. Considering the opportunity cost, this may be less profitable than investing in other directions.
Second, this is still under the expectation that computing power demand will continue to grow massively. What if AI development is blocked and expansion stagnates? What if the number of players decreases and the current training computing power usage, which accounts for more than 50%, drops sharply? What if major breakthroughs in infra reduce computing power demand? These risk factors will obviously further weaken the expected stable return of such large-scale upfront investment.
Simply put, even if we calculate the accounts under the best current scenario, the profit may not be considerable. The risk is too high, and the maximum return cannot be obtained solely through computing power expansion. Only one player can win the game, and if other players can only participate as computing power providers, their upper limit of return is not high, while the lower limit is very low.
Credit rating agencies are also using this set of criteria to examine the hematopoietic bottom line of each giant. S&P Global maintained Meta's AA− credit rating on October 30, but downgraded its rating outlook to negative. The core reason is precisely the concern that aggressive AI capital expenditure will erode cash flow, which may cause Meta's leverage ratio to break through the 1.0x upper limit allowed by the AA− rating in early 2027.
Similarly, we can understand why Anthropic was rated as junk credit.
On October 30, the Financial Times and multiple market institutions disclosed that Blackstone, in communication with investors, predicted that even after Anthropic completes its listing, its credit rating may remain in the non-investment grade (commonly known as the "high-yield" or the high-end range of junk grade), and it is difficult to obtain an investment-grade rating in the short term. This is also because Anthropic's IPO prospectus (Form S-1) publicly submitted on October 26 shows that the company is burdened with high-value equipment lease commitments for up to five years (the total five-year lease related to Broadcom alone reaches $161.2 billion, excluding interest), and the revolving credit agreement established during this period further aggravates this concern.
Since Anthropic has no positive free cash flow, it can only rely on borrowing to build its computing power infrastructure. And this borrowing requires collateral and underwriting guarantees.
On October 26, Broadcom, together with Bank of America, Blackstone, Citigroup, and Morgan Stanley, announced a $630 billion equipment financing arrangement to deploy exclusive AI computing power and network equipment based on Broadcom's technology for Anthropic, with delivery expected to start in 2027.
Chip seller Broadcom personally acted as the underwriter for the priority lender. In the 8-K document submitted to regulators, Broadcom promised to provide residual value guarantee of up to approximately $43 billion (undiscounted caliber) for the priority loans. In the event of a default and the proceeds from the sale of the cabinets are less than 72.5% of the outstanding lease principal, Broadcom must pay the difference out of its own pocket.
Similar revolving loans further intensify the transmission risk of infrastructure vulnerability. If Anthropic fails, Broadcom will also suffer.
Doubts about long-term payment ability soon spread to the upstream hardware supply chain. At 09:45 a.m. ET on October 30, the storage chip and hard disk sectors were collectively sold off. Intraday records showed Micron fell 4.26%, Sandisk plunged 8.53%, Western Digital fell 6.07%, and Seagate fell 5.26%. It is intriguing that right after the market closed the previous day, Sandisk had just delivered an extremely impressive fiscal 2027 Q1 earnings report: revenue for the quarter reached $11.031 billion, up 23% quarter-over-quarter and a surge of 378% year-over-year (of which data center business revenue increased 61% quarter-over-quarter to $4.802 billion). GAAP gross margin jumped from 29.8% a year ago to 85.0%, and the company also announced an 8-for-1 stock split plan and raised its next quarter's revenue guidance to $12.2 billion to $12.6 billion. Strong performance not only failed to push up the stock price, but also triggered profit-taking and a resonant decline in the sector.
Because everyone is predicting that this crazy and aggressive infrastructure investment may not be sustainable. And the future expectations of hardware suppliers that have been advancing rapidly as a result will naturally be adjusted downwards accordingly.
AI scientific exploration ushers in an explosive period, crown research in various fields is broken through
In late October 2026, the role of artificial intelligence in the field of basic science underwent a watershed qualitative change. Within just a few days, cutting-edge laboratories such as OpenAI and Anthropic intensively disclosed a number of breakthroughs spanning theoretical computer science, basic mathematics, theoretical physics, chemical kinetics, and biomedicine.
Different from early AI which was only limited to local engineering tuning under given evaluation indicators, this round of achievements directly pierces through the core conjectures and experimental black boxes that have plagued the human academic community for decades. From the two institutions independently overthrowing the computational complexity cornerstone "Exponential Time Hypothesis (ETH)" on the same day, to fully conquering the local smoothing conjecture and the Kakeya conjecture in harmonic analysis.
From the 1,147-page construction of the existence of four-dimensional quantum field theory, to the cross-model relay completing the formal proof of the global convergence of chemical reaction networks, and then going deep into the automated wet laboratory to overthrow the wrong scoring standards of traditional instruments.
A complete scientific research chain extending from "formal axiom deduction" to "real physical world experimental design" is being rapidly opened up.
On October 26, OpenAI dropped a blockbuster in its official announcement Working with academic communities to build scientific understanding, publicly releasing a 67-page preprint paper titled The Exponential Time Hypothesis is False and supporting Lean 4 formal proof code, announcing that its internal reasoning model found a natural language proof to overthrow the "Exponential Time Hypothesis (ETH)" on October 16, and completed machine formal verification the next day.
Dramatically, just one day later on October 27, Anthropic also published the paper ATime Algorithm for 3-SAT on the mathematical preprint platform Hexagon and its own formal mathematical repository. Anthropic's attribution statement shows that its internal research version of the Claude model also independently discovered the core algorithm to overthrow ETH on October 16, consuming 14 million output tokens without any human intervention, and then completed the Lean formalization after human researchers assisted in improving the presentation.
While announcing the computational complexity results, OpenAI also released another fully public natural language paper and Lean proof package Local smoothing for the Euclidean wave equation to the pure mathematics community in the same October 26 announcement (the internal model produced the natural language proof on October 18, and completed the Lean formalization on October 19). This work directly claims to have solved the "Local Smoothing Conjecture" of the Euclidean wave equation proposed by harmonic analysis master Christopher Sogge in 1991, and incidentally advanced the famous Kakeya conjecture, Nikodym maximal conjecture, and Stein Fourier restriction conjecture in geometric measure theory and harmonic analysis to any dimension.
In the history of human mathematicians' advancement, every upward dimension is accompanied by extremely difficult geometry and analysis breakthroughs: Larry Guth, Hong Wang and Ruixiang Zhang, in their long paper submitted in 2019 and published in Annals of Mathematics in 2020, first conquered the local smoothing conjecture in two-dimensional space. In 2025, Hong Wang and Joshua Zahl jointly proved the Kakeya conjecture in three-dimensional space.
In addition to the two open-source code achievements in mathematics and computer science, the third breakthrough announced by OpenAI in the October 26 announcement points to one of the holy grails of modern theoretical physics. Its internal model generated a natural language proof of the mathematical existence of four-dimensional pure Yang–Mills gauge field theory on September 9, further generated a four-dimensional interacting quantum field theory including gauge fields and fermions (matter particles) on September 18, including the existence proof of quantum chromodynamics (QCD) describing strong interaction and the unified completeness scheme of the particle physics standard model, and completed the Lean formalization of the pure Yang-Mills existence part on October 16.
Different from the previous two practices of directly putting them online for anyone to download, this batch of physical materials has not been made public for the time being, but has been fully handed over to a five-person independent scholar working group chaired by 2004 Nobel Prize in Physics laureate David Gross, composed of Arthur Jaffe, Michael Douglas, Igor Klebanov and Nikita Nekrasov. The working group will decide when and how to disclose it to the academic community.
According to the official statement of the five-person working group released by the Kavli Institute for Theoretical Physics (KITP) on October 28, the natural language proof manuscript generated by the OpenAI model is as astonishingly long as 1,147 pages. From the preliminary information disclosed by the working group, AI did not invent a brand new set of physics from scratch, but pushed all the way along the "Renormalization Group" construction route pioneered by Polish mathematical physicist Tadeusz Bałaban in the 1980s. Due to the "Asymptotic Freedom" property discovered by David Gross et al. in 1973, the strong interaction becomes very weak at very short distances (high-energy ultraviolet region). Bałaban took advantage of this point back then and wrote more than a dozen extremely obscure long articles that almost no peers could fully read, trying to control the ultraviolet divergence when the grid is refined step by step. The working group pointed out that OpenAI's model drew on and completed Bałaban's core idea, tracking the renormalization group analysis all the way from the microscopic lattice to infinite distance, thus claiming to prove the existence of the appropriate continuous limit.
If we shift our attention from the three major theoretical breakthroughs above to a wider range of scientific daily life, we can see that AI is not only conquering top-level problems at a single point, but also reshaping the entire production pipeline of scientific research in the way of "cross-model division of labor, reusing open source formal assets, and deeply intervening in real experimental design".
On October 27, the Anthropic Life Sciences team released a progress report on drug R&D cooperation, systematically disclosing the application results of Claude in the real R&D pipelines of six biomedical institutions (BigHat Biosciences, Generate Biomedicines, Divergence Labs, Enveda, The Column Group and Bristol Myers Squibb) in the past few months, and simultaneously announced the acquisition of Mithril Medicine, a rare disease R&D company with only one employee.
In these cases, the role of AI is no longer limited to the early "molecular structure generator": at Generate Biomedicines, scientists used Claude Code to shorten the intracellular signaling domain with unclear structure in CAR-T cell therapy by about 70%, increasing the receptor expression on the cell surface by about 10 times, freeing up valuable gene vector space for conquering solid tumors. At Divergence Labs, Claude independently planned and generated about 35,000 bispecific antibody designs within one month, successfully screening out candidate molecules that can strip bound allergen antibodies (IgE) from immune cells at conventional injection concentrations. And the most noteworthy breakthrough in scientific methodology occurred in the antibody lead optimization cooperation with BigHat Biosciences.
By comparing the original data, the researchers discovered the limit problem of the detection equipment. To verify this conjecture, he asked Claude to build a biophysical model explaining the trend of the original curve, and voluntarily proposed to modify the measurement protocol of the wet experiment. BigHat's scientists perfected and carried out this set of controlled experiments, and the results fully confirmed Claude's physical deduction. The team immediately abolished the original wrong scoring formula, and used the newly revised evaluation indicators to guide subsequent iterations. The final optimal antibody variant not only increased the target binding affinity by about 9 to 10 times (reaching sub-nanomolar affinity), but also increased the tolerance temperature of the anti-cancer lead compound to maintain a dissolved state by 18°C, and the aggregation rate of the long-acting drug lead compound slowed down by more than 100 times. Some molecules did not even fully aggregate under the maximum 85°C limit test of the instrument.
This case pushed the boundary of AI participating in scientific research a big step forward. In most of the past "AI pharmaceutical" narratives, humans are responsible for setting the exam paper and scoring formula, and the model is just a tireless score brushing machine, responsible for finding the one with the highest score faster among millions of candidate molecules. But this example shows that AI has the ability to provide more fundamental underlying assumptions under human guidance.
These cases reveal the most profound changes in current AI empowering science. AI has grown from the "laboratory apprentice" who only used to do superficial parameter tuning in the past, to a discovery engine that can produce thousand-page-level underlying constructions and challenge the core assumptions of the field.
Google announces Gemini 4 Carbon, temporarily taking the lead
Only two weeks after the release of the previous generation cutting-edge model Gemini 4 Argon, Google officially announced the new flagship model Gemini 4 Carbon on October 28.
To prove Carbon's capability leap, Google released a number of third-party and internal hardcore evaluation data. In the newly updated independent benchmark Artificial Analysis Intelligence Index v5.0, Carbon topped the list with a score of 59; on the front-end and full-stack development ranking platform WebDev Arena, its Elo score jumped 89 points sharply compared to the previous generation, ranking fourth with 1768 points. More iconic is its performance in the full life cycle of cybersecurity: in the SEC-bench Pro test for mining real software vulnerabilities, Carbon achieved an accuracy of 88.2%, got 54.8% in the ExploitGym benchmark for verifying whether vulnerabilities are truly exploitable, and reached 71.0% on CWE-bench, which directly converts vulnerabilities into code patches.
Google's internal team observed in the