ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-0956 posts

Refresh

HOPE: Incorporating Expert Pairwise Interactions into Quadratic Programming to Maintain MoE Cooperative Groups at 50% Pruning Rate

Topic · MoE专家协作剪枝Material
2026-11-09 20:34 GMT+8

The vast majority of parameters in Mixture-of-Experts (MoE) models remain idle—of the approximately 250 GB of VRAM occupied by Qwen3.5-122B, 95% is taken up by the MLPs of experts across all layers, yet only 8 experts are actually activated per token. Directly deleting inactive experts is the most concise path to compressing this memory bottleneck: it requires no training, no dynamic loading, and can be combined with quantization. The difficulty lies in deciding whom to delete. Existing pruning methods—such as REAP, MAN, and EAN—assign a scalar importance score to each expert independently and then remove those with the lowest scores. Author Alex M. Tseng (Amazon) points out that this approach assumes each expert's contribution is independent of others, but experimental observations indicate that experts in MoEs form stable cooperative groups that are consistently co-activated. Independent scoring completely ignores this: removing expert A may seem harmless on its own, but if A frequently cooperates with high-importance expert B, a single deletion could dismantle an entire cooperative group.

The core of HOPE is an E×E interaction matrix F. The diagonal of the matrix measures the independent importance of individual experts—which correlates highly with REAP rankings, with a Spearman correlation coefficient of 0.988—while the off-diagonal elements capture the average gate-weighted output product of experts i and j on tokens where they are co-activated. With this matrix, selecting the pruning set becomes a quadratic programming problem: minimizing the total interaction strength among pruned experts. The solution employs continuous relaxation followed by rounding, taking 1 to 2 seconds per layer, which is negligible compared to the total overhead of pre-calibration forward propagation.

The advantage of this mechanism is most significant at a 50% pruning rate. Taking Qwen3.5-35B-A3B as an example, after HOPE pruning, the model scored 42.3 on SWE-Bench Pro, matching the unpruned model's 41.8; whereas REAP dropped to 36.2 under the same conditions, representing a regression of about 13%. The authors attribute this to the fact that at high pruning rates, many cooperative links face the risk of being severed, while HOPE's quadratic programming automatically sacrifices some experts with high independent scores but weak cooperative node status to preserve the overall cooperative structure.

However, the authors also acknowledge the boundaries of HOPE: at low pruning rates of 10% to 20%, first-order methods are sometimes equally competitive, narrowing HOPE's advantage; the current implementation prunes each layer independently without considering cross-layer interactions; the modeled relationships are limited to pairs, making triplet and higher-order interactions combinatorially difficult to estimate. All performance figures are self-tested by the author's team and have not yet been independently reproduced.

Sources:huggingface.co

Research

Tesla CEO Musk Says Cybertruck Orders Exceed Capacity, Mass Deposit Refunds Coming Before Shutdown

Material
2026-11-09 18:13 GMT+8

Tesla CEO Elon Musk stated on November 9, 2026, that the number of Cybertruck orders "far exceeds" Tesla's capacity to build them before the electric pickup is discontinued next year. Writing on X, Musk said, "We will try to make as many as possible before shutting down the line in about 8 months, but will have to return deposits for a lot of people."

This statement adds specific operational detail to previous announcements regarding the wind-down of Cybertruck production. Musk emphasized that the decision was not driven by unprofitability, but by the priority of scaling Tesla's Cybercab robotaxi. He noted that the opportunity cost of maintaining the complex Cybertruck supply chain and engineering resources vastly outweighs the value of incremental Cybercab production.

For investors, the confirmation of mass deposit refunds implies immediate short-term cash flow obligations and pressure on customer relations. While the exact number of affected customers and total refund value remains undisclosed, this fact underscores Tesla's accelerated pivot from traditional passenger vehicles to AI-driven robotics and mobility. This follows earlier decisions to discontinue the Model S and X to free up factory space for Optimus humanoid robots.

Sources:Business Insider

Research

Semafor Business & Finance Editor Liz Hoffman: AI Trade Unwinding Will Reach Beyond Silicon Valley to Glass, HVAC and Other Sectors

Topic · AI泡沫退潮冲击Material
2026-11-09 21:27 GMT+8

In a column published on November 9, Semafor Business & Finance editor Liz Hoffman posited a key judgment: the AI trade is unwinding, and its impact will be broader than what markets are currently pricing in. Her core argument rests not on how much any single stock has fallen, but on a structural fact—that this wave of AI enthusiasm has drawn a vast number of industrial companies outside Silicon Valley into the same narrative framework, making them unable to escape when the tide turns.

Hoffman uses two threads to illustrate changes in the funding environment. OpenAI originally planned to raise $30 billion, money it was reportedly able to assemble 'with its eyes closed' just a few months ago, but which now appears likely to shrink significantly. Meanwhile, SpaceX, an investment-grade giant, is being priced in the debt markets like a junk-rated startup. These events alone do not signal the end of the AI era, but Hoffman argues they mark the evaporation of the underlying psychology driving these transactions—the combination of FOMO (fear of missing out) and the 'greater fool' theory of investing.

What adds incremental value to this column is her tracking of 'spillover effects.' Corning’s stock doubled in the first half of 2026 following a $6 billion data center supply agreement with Meta; by early November, almost all those gains had been given back. Modine, formerly a car radiator manufacturer, announced in January that it would spin off its traditional business to become a 'pure-play climate solutions company focused on the high-growth data center market,' yet its stock now trades lower than before that announcement. Hoffman also notes that paint supplier Sherwin-Williams and HVAC manufacturer Carrier are in the midst of their own pivots toward AI.

Her conclusion is that as the market returns to doing its job—distinguishing winners from losers—some of the losers will be household names with twelve-figure valuations. And because the AI boom’s reach extended into glass, HVAC, and paint industries, the unwind will extend there as well.

It is important to note that the 'spark' examples Hoffman lists—data center delays triggering guarantee disputes among hyperscalers, swarms of rogue agents hacking into a major city’s electric grid, or Nvidia CEO Jensen Huang admitting to an affair with a suspected Chinese spy—are rhetorical possibilities, not predictive models. The column provides no quantitative framework to define the threshold or duration of an 'unwind.' Her argument is directional: funding conditions are tightening, and the AI re-evaluation of non-tech companies is already reversing; there is a causal link between these two factors. Whether this constitutes systemic contagion or merely independent company-specific issues remains a boundary for readers to judge for themselves.

Sources:semafor.com

Research

William Imoh, Founder Who Never Wrote Swift, Builds Native Voice App with Codex: 1,000 Days on Duolingo Didn't Teach Him Norwegian

Topic · Codex原生应用开发Material
2026-11-09 21:04 GMT+8

After moving to Norway, William Imoh spent 1,000 days studying Norwegian via Duolingo but still found himself unable to conduct a full conversation. This specific frustration directly led to the creation of Mural, a voice-learning application he later built using Codex.

His background does not point to that of a mobile developer. He holds a degree in chemical engineering, taught himself to code, worked as a product manager for four years, and had never written a single line of Swift. However, in an episode of Lenny's Newsletter's How I AI podcast, he demonstrated how he used Codex and GPT-Live to build native applications for both iOS and Android. Mural is a real-time voice application whose core function allows learners to practice conversations in Norwegian.

During the podcast, he discussed the importance of design language—a specific design prompt made Mural’s appearance distinct from other language-learning apps on the market. He also addressed the per-session cost of GPT-Live-1 and latency thresholds, factors that determine whether a voice-tutoring experience is viable. While these specific figures were detailed in the show, this summary confirms only that they were discussed.

On the operational side, William Imoh organized his team of Grok Bot agents by assigning roles based on characters from The Lord of the Rings. He mapped the various responsibilities of a solo founder onto different characters, using this system to assist in managing a team of about 25 people. The episode also noted that he granted his agents access to 15 open tabs, with one exception.

He mentioned that even after model improvements, there are still certain tasks he must complete personally. What keeps him awake at night is not the fear of being replaced, but rather the fear of dependency.

The practical boundaries of this case are significant: Mural is a voice-conversation application with a relatively clear task structure (inputting voice, outputting voice, and evaluating conversation quality), and Imoh himself is the target user. The conditions that allowed a product manager who had never written Swift to complete this project may not be directly transferable to applications requiring complex state management, backend architecture, or large-scale user testing. His experience illustrates the actual capabilities of current AI coding tools for specific types of tasks, rather than representing a universal replacement for software development.

Sources:lennysnewsletter.com

Research

German ifo Institute Releases AI Early Indicator: Entry-Level Job Decline Associated with AI, Youth Employment Down 14% in High-Exposure Roles

Topic · ifo AI劳动力指标Material
2026-11-09 21:30 GMT+8

The ifo Institute in Munich published a new AI early indicator (KI-Frühindikator) on November 9, 2026, designed to measure the impact of widespread AI adoption on the labor market. The study compared employment and new hires across 140 occupations, relating them to the substitutability of human work by AI.

According to the data, full-time employment among 20- to 24-year-olds in the most AI-exposed occupations has fallen by 14% since 2022, dropping from approximately 279,000 to 239,000. These high-exposure roles include data entry and text processing, software developers, accountants, database and network technicians, as well as authors and journalists. "We do not yet see broad job cuts due to AI," said Clemens Fuest, President of the ifo Institute. "But those starting their careers feel the change first."

However, the impact of AI on the labor market is not uniformly negative. The study notes that while entry-level positions have decreased, total employment in AI-exposed occupations actually grew by 1.8% since 2022, whereas employment in non-exposed occupations declined by 6.4%. This suggests that AI is altering the nature of work through 'augmentation' rather than just 'automation,' enhancing the productivity of existing employees. Co-author Oliver Falck views the loss of jobs for beginners as an early warning sign for the labor market.

It is important to note that researchers emphasize this indicator does not provide a monocausal explanation. Demographic changes and longer training periods may also contribute to the observed employment trends. Additionally, even executives face risks, as AI can handle significant portions of management tasks; conversely, occupations relying primarily on manual labor, such as construction workers, carpenters, cleaners, and farmers, are least affected by AI.

Sources:t3n.de

Research

Open-Source Project Darling Adopts AI-Assisted Contributions, Runs League of Legends on Linux

Topic · DarlingQuick take
2026-11-09 21:02 GMT+8

The macOS compatibility layer project Darling has officially adopted rules for AI-assisted code contributions. Designed to allow Linux users to run macOS binaries, the project had progressed slowly for over a decade due to limited developer resources. The team states that large language models (LLMs) have changed the equation, making previously unsolvable problems financially and practically realistic.

According to Phoronix, the Darling team used AI-generated patches to successfully run the macOS version of League of Legends on Linux. This breakthrough demonstrates that for small open-source teams, AI tools can significantly lower the barrier to complex compatibility development, allowing them to compete with Apple's ecosystem on closer to level ground.

The policy explicitly views AI as a powerful tool but emphasizes that its use must be bound by human oversight. Currently, several Darling forks embracing AI-generated code are working to merge these patches into the upstream repository.

Sources:phoronix.com

Research

Korean Creator Solutions Firm Adoba Launches AI Content Platform with Tencent Cloud, Enabling Comparison of Up to 12 Models

Topic · Adoba腾讯云AI平台Quick take
2026-11-09 07:58 GMT+8

Global creator solutions company Adoba has announced the launch of 'adobaRo AIGC Studio', an integrated AI content creation service developed in technical cooperation with Tencent Cloud. The platform aims to streamline the fragmented workflow where creators often switch between multiple AI services to generate and verify video and image content.

The core feature, dubbed 'AI Battle', allows users to input a single prompt and view results from up to 12 different generative AI models side-by-side. This design addresses the challenge of selecting the most suitable model for specific creative tasks, as different models excel in different areas, thereby reducing the burden of choice for creators.

Adoba stated that the service leverages its experience in helping over 1,500 creator teams enter the Chinese market and operating more than 2,500 global channels, combined with Tencent Cloud's AI technology. An Jun-han, CEO of Adoba, noted that choosing among numerous AI models has become a new challenge for creators, and the platform was designed to reduce this complexity. Heo Jung-pil, Country Manager of Tencent Cloud Korea, emphasized the significance of connecting diverse generative AI technologies into a practical service for creators.

Sources:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

Research

Google Launches Two New TPU Variants as External Customer Count Nears 100

Material
2026-11-09 20:59 GMT+8

On November 9, 2026, Google announced the addition of two new systems to its eighth-generation AI accelerator lineup: the TPU 8tx for confidential computing and the TPU 8i Max for memory-intensive workloads. This marks the first time Google has shipped four TPU systems within a single year, signaling a strategic shift from treating TPUs as an internal tool to marketing them as a competitive platform for external customers.

The TPU 8tx targets enterprises and government agencies working with confidential data. While data is typically encrypted during storage and transmission but decrypted in memory during processing, Google states that the TPU 8tx keeps data and model weights protected and isolated even during active use. The system uses the low-latency Boardfly network and delivers up to twice the latency reduction compared to the TPU 8t.

The TPU 8i Max addresses the memory bottleneck common in AI agent applications handling large codebases or video. By placing two TPU 8i chips on a single package, Google achieves double the compute power, double the memory, and double the memory bandwidth within the same rack format, significantly increasing performance per square meter of data center space.

Beyond hardware, Google released TorchTPU as open source to allow PyTorch code to run on TPUs with minimal changes, alongside support for inference frameworks like vLLM and SGLang. With nearly 100 external TPU customers, Google is increasingly positioning its in-house chips as an alternative to Nvidia GPUs, particularly for day-to-day model inference.

Sources:The Decoder

Research

Investigation: Wearable Sleep Trackers Using Brainwaves to Train AI Models Without Explicit Consent

Topic · 可穿戴脑电AI训练争议Material
2026-11-09 20:00 GMT+8

According to a report by Tom's Hardware, a recent investigation by Science magazine and the Pulitzer Center's AI Accountability Network found that several consumer neurotechnology companies are using users' electroencephalography (EEG) data—electrical signals generated by brain activity—to train artificial intelligence models, without having previously disclosed this use clearly in their privacy policies. These companies have accumulated tens of millions of hours of EEG recordings.

Canadian firm Interaxon, maker of the Muse sleep and meditation headband, is one of the named entities. Its devices monitor electrical activity in the brain alongside other physiological measurements while users sleep. Nadia Kumentas, Interaxon's Chief Marketing Officer, stated that the company has already trained an AI model using about 50,000 hours of EEG recordings and claims it removes identifying information such as names, emails, and addresses before using the data. However, prior to speaking with Science, Interaxon's privacy policy did not explicitly disclose that users' brain recordings could be used to train foundation models. Kumentas acknowledged, "We agree users should be explicitly informed." For customers whose recordings had already been used, this classification came late, likely exposing the company to potential lawsuits.

Emotiv, another wearable EEG maker, is also training its Large Brain model partly on recordings from customers using its high-end devices. Following questions from Science, Emotiv updated its privacy policy to disclose the use of de-identified neural data for model training subject to opt-in consent.

The logic behind these systems differs from training large language models. Instead of feeding AI enormous amounts of text, developers supply it with recordings of electrical activity in the brain. The model then learns recurring patterns that can be adapted to recognize different brain states or assist with specific research tasks. For instance, AI startup Piramidal trained its model on more than 4 million hours of EEG data in partnership with Cleveland Clinic, developing a seizure-detection system that correctly identified nine out of ten seizures in a pilot study. However, this does not mean consumer headbands can read thoughts; most wearable EEG devices use only one to six electrodes, considerably fewer than the 19 to 25 typically found in clinical EEG systems, and their recordings are generally less comprehensive.

Privacy implications in wearable brain technology are not entirely new. In February, an engineer discovered he could access other people's brainwave recordings through a smart sleep mask he had purchased. Furthermore, the investigation revealed that more than two-thirds of the companies studied explicitly allowed corporate transfers of data without user consent. This could be significant if a small neurotechnology firm is acquired, potentially putting its accumulated recordings in the hands of a new owner. Currently, Colorado and California in the United States have amended their consumer privacy laws to cover neural data, although other states lag behind.

Sources:Tom's Hardware

Research

Score Centering: Eliminating Gradient Drift in RL Training with Additive Correction

Material
2026-11-09 18:33 GMT+8

Reinforcement learning (RL) training for large language models (LLMs) typically employs a disaggregated architecture: an inference engine is responsible for generating rollouts, while a training engine computes policy gradients. These components run on different hardware, utilizing distinct precisions, kernels, and batch shapes; consequently, they produce different probabilities for the same token—a phenomenon known as Training-Inference Mismatch (TIM). Asynchronous training further exacerbates this issue, as policy updates may be based on rollouts that are already stale.

The core contribution of Marek and Ryabinin’s paper lies not merely in proposing a correction method, but in providing a deep analysis of why TIM causes instability. They demonstrate that under a constant reward setting, the theoretical policy gradient should be zero (since all actions receive the same reward, there is no learning signal). However, when TIM is introduced, the probabilities of the sampler q and the trainer p no longer cancel out, causing the expected gradient to become non-zero—the trainer begins to mimic the distributional differences of the sampler regardless of whether the reward provides meaningful information. This phenomenon is referred to as gradient drift.

More generally, by decomposing the policy gradient into covariance terms, two parts can be separated: the drift term equals the expected reward multiplied by the expected score of the ratio of sampler to trainer probabilities; the signal term is the covariance between the reward and the score. The drift term does not distinguish between successful and unsuccessful rollouts, relying solely on the mean of the rewards. In online RL, where the sampler is continuously refreshed from the trainer, drift forms a feedback loop: TIM leads to biased updates, which propagate to the sampler, which then produces new biased rollouts—errors accumulate gradually until training collapses.

Score Centering works by subtracting the expected score under the sampler from the token-level scores calculated by the trainer at each prefix position. This subtraction ensures that the centered scores have an expectation of zero under the sampler distribution, thereby completely eliminating the drift term in the covariance decomposition while leaving the signal term unchanged (since subtracting a constant does not affect covariance). Unlike importance sampling, this is an additive correction that avoids high-variance multiplicative weights and requires no clipping or masking to limit extreme values.

In practical implementation, precisely calculating the expected score over the entire vocabulary for the sampler is prohibitively expensive (LLM vocabularies typically exceed 150,000 tokens). The paper employs a top-k approximation: retaining only the 128 tokens with the highest sampler probabilities and approximating the remainder using scaled tail probabilities from the trainer. A key derivation utilizes the property that the expected score is zero under the trainer’s own distribution, allowing the correction term to be computed using only top-k tokens. Experiments show that k=128, and even k=32, achieve effects comparable to full-vocabulary approximation, though the guarantee of drift elimination is only approximate in these implementations.

Sources:cameronrwolfe.substack.com

Research

US Power Firm NRG Energy Signs Agreement to Sell Six Gas Plants to LS Power for $4 Billion to Fund Buybacks and Debt Reduction

Topic · NRG EnergyMaterial
Verified 2026-11-09 20:33 GMT+8

NRG Energy, Inc. (NYSE: NRG) announced on November 9, 2026, via an SEC Form 8-K filing, that it has entered into a definitive agreement to sell six natural gas-fired generation assets to an affiliate of LS Power. The purchase price is $4 billion in cash, subject to customary working capital and other closing adjustments. The transaction implies a 2027 EV/Adjusted EBITDA multiple of 8.0x.

The assets represent 4.9 gigawatts (GW) of generation capacity across the New York Independent System Operator (NYISO), ISO New England (ISO-NE), and Electric Reliability Council of Texas (ERCOT). Specific plants include Ravenswood, Ocean State Power, Wallingford, Jack County, Johnson County, and R.W. Miller. Upon completion, NRG will own and operate 20 GW of generation capacity nationwide, down from approximately 25 GW, with 10 GW each in ERCOT and PJM.

Robert Gaudette, President and CEO of NRG, stated that the transaction unlocks substantial value while maintaining a robust generation footprint. Proceeds will support NRG's 2027 pro forma capital allocation plan, including $3 billion in share repurchases across 2026 and 2027, with incremental buybacks expected to begin before the end of Q4 2026. Additionally, NRG plans an extra $2 billion in debt reduction upon closing, bringing total planned 2027 debt reduction to $3 billion to maintain current leverage profiles. Remaining proceeds are reserved for contracted growth investments or further debt reduction.

The sale is targeted to close in the second quarter of 2027, subject to customary conditions and regulatory approvals, including Hart-Scott-Rodino (HSR), Federal Energy Regulatory Commission (FERC), and New York State Public Service Commission (NYSPSC). Barclays serves as lead financial advisor, and White & Case LLP acts as legal counsel.

Sources:sec.gov

Research

Ilman Shazhaev, Founder of Abu Dhabi AI Company Dizzaract: The True Test of a Unicorn Is Renewal at Full Price, Not Paper Valuation

Topic · DizzaractMaterial
2026-11-09 20:00 GMT+8

In the first half of 2026, 250 companies joined the Crunchbase unicorn list, making a $1 billion valuation seem increasingly attainable. But Ilman Shazhaev, founder and CEO of the Abu Dhabi-based AI and gaming infrastructure company Dizzaract, proposed a different benchmark: what truly matters is not the valuation, but whether the company can maintain $1 billion in sustained revenue from customers who renew at full price.

His reasoning begins with the structural changes in AI performance pricing. Citing research from Epoch AI, Shazhaev notes that since 2023, the cost to achieve a fixed level of AI performance has dropped thirteenfold every year. This means companies that simply resell model access face brutal price competition—OpenRouter’s $113 million Series B illustrates that developers can already switch between different models through a shared interface. In such an environment, ARR figures alone no longer indicate defensibility.

Shazhaev believes that what is truly hard to replace is not the underlying model, but software embedded in customer approval processes, records, and daily decision-making. His test method is specific: Do customers renew at full price? Are contracts expanding? Do margins remain healthy after deducting inference costs? These three questions reduce 'revenue' from vague annualized run rates to trackable business facts.

The most actionable part of the article lies in its discussion of capital structure. Shazhaev advocates that founders cross borders when raising funds, because local investors can simultaneously serve as market entry points. He cites the $30 million Series A for London- and Gulf-based 1001 AI as an example: U.S. venture firm Lux Capital led the investment, while Sanabil Investments, backed by Saudi Arabia's Public Investment Fund (PIF), participated. Engineers are in London, and customers are in the Gulf. While this deal itself cannot prove future retention rates, it demonstrates how capital, talent, and customers can be distributed across different maps.

More critically, he distinguishes between the differing obligations of two types of capital. Traditional venture capital and sovereign capital are often simplified into a dichotomy of 'fast money' versus 'patient capital,' but Shazhaev argues that a more practical distinction lies in the obligations attached to each check. Venture funds need to generate returns within their fund's lifespan, so they focus on commercial momentum and exit scale; sovereign capital usually asks where user data is processed, which technical capabilities remain domestic, and whether local customers can continue to rely on the service years later. Some sovereign funds even treat local development missions as evaluation criteria equal to financial returns.

He simultaneously provides important caveats: these models are not universally applicable. Many sovereign funds pursue conventional financial returns, traditional VCs can support long-cycle enterprise sales, and government-affiliated capital does not guarantee procurement contracts. Founders' actual work involves reading the fine print—governance rights, exclusivity requirements, committed commercial introductions—and then judging whether the investor's regional reach is worth those constraints.

Sources:Crunchbase News

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

Under mild TIM caused by quantization, Score Centering alone is sufficient to stabilize training. However, under severe staleness (such as in asynchronous training where the sampler synchronizes only once every 64 steps), Score Centering alone is insufficient—because the learning signal is still measured under the sampler distribution, the deviation from the trainer distribution requires correction via importance sampling. The paper suggests combining Score Centering with truncated or masked importance sampling when significant policy differences may arise between the trainer and the sampler.

Cameron Wolfe is a Deep Learning Researcher at Netflix holding a Ph.D., known for his communications that deconstruct AI research papers. He cites Nathan Lambert, noting that many frontier RL laboratories are already using this method, albeit previously unpublished.

Shazhaev concludes that the best capital strategy follows the customer path, from available product to renewal. Venture capital accelerates development and sales, provided the business model is already proven; regional strategic investors truly add value only when new markets' infrastructure and procurement rules require local expertise. He explicitly states that he would rather show investors a small batch of customers renewing at full price than present a glossy revenue forecast dependent on the next round of financing. The numbers that matter are net revenue retention, production deployment counts, and profit margins after delivery costs.

The limitation of this framework lies in the author's position of interest. Dizzaract itself is an Abu Dhabi-headquartered company, the largest game development studio in the Middle East, and an AI infrastructure provider; describing sovereign capital's localization obligations as advantages rather than constraints aligns directionally with its own financing needs. Additionally, the citation of a 13-fold decline in performance costs comes from third-party research, and the author did not independently verify the applicability of this trend to his specific sub-market. However, as a checklist for evaluating the capital structure and defensibility of AI companies, the three dimensions of full-price renewal, post-delivery profit margins, and obligation matching possess operational specificity.