# Сопутствующие статьи по теме AI

Новостной центр HTX предлагает последние статьи и углубленный анализ по "AI", охватывающие рыночные тренды, новости проектов, развитие технологий и политику регулирования в криптоиндустрии.

When Tokens Cost More Than People, 'AI Narrative' Runs Into Trouble

Title: When Tokens Cost More Than People, the "AI Narrative" Hits Trouble The economic sustainability of corporate AI adoption is under scrutiny as token consumption soars while measurable business value remains elusive. Major companies like Uber and Microsoft report struggling to justify rising AI costs, with executives coining terms like "tokenmaxxing" to describe wasteful usage. Data reveals a stark picture: for every dollar spent on AI tokens, only 18 cents translates to user-facing value, with the rest consumed by bug fixes, rework, and friction. The debate splits into bullish and bearish camps. Bulls, like Goldman Sachs analysts, see current inefficiencies as growing pains, predicting a 24-fold increase in token demand by 2030 and a shift towards healthier metrics like "cost per effective action." They point to indicators of real productivity gains and argue current tech valuations are not in bubble territory. Bears, however, highlight an unsustainable model where value is heavily concentrated in semiconductor companies like Nvidia, funded by cloud giants taking on massive debt. Studies show 95% of firms investing in generative AI see zero return. A deeper concern is the circular financial structure between cloud providers (hyperscalers) and AI labs like OpenAI and Anthropic. Billions in cloud service commitments are tied to these labs, which are partly funded by the hyperscalers' own investment. This creates a loop where cloud revenue depends on labs securing continuous external funding to pay their compute bills, which in turn relies on end-corporates willing to pay ever-higher token costs. The sustainability of this cycle is now in question. While not a classic bubble—AI technology is real and delivers productivity for power users—the central issue has shifted. The focus is no longer just on technological capability but on economics: whether the savings AI generates for businesses can outpace the soaring costs and justify the valuations of labs and cloud providers. The era of equating rising token usage with successful AI transformation is over. The bill for AI has arrived, but who ultimately pays remains uncertain.

marsbit05/29 01:44

When Tokens Cost More Than People, 'AI Narrative' Runs Into Trouble

marsbit05/29 01:44

Li Kaifu and Wang Xiaochuan Pivot: The First Half of the Large Model Entrepreneurship Era Ends

Li Kaifu and Wang Xiaochuan, leading figures in China's AI industry, are signaling a strategic shift, marking the end of the first phase of the large language model (LLM) startup boom. Li's 01.AI, once seen as a potential "Chinese OpenAI," is now pivoting towards enterprise applications and Agent technology, explicitly modeling itself after the低调但 profitable Palantir with a goal of profitability by 2026. Wang's Baichuan Intelligence is fully转战ing the vertical field of healthcare, launching a medical LLM and AI doctor product. This reflects a broader industry清醒. The initial狂热 of 2023, with its focus on chasing参数, benchmarks, and the "Chinese OpenAI" narrative, has collided with the harsh reality of an AI "heavy industry" war dominated by immense capital expenditure from US tech giants (微软, Google, etc.) and Chinese互联网大厂. The cost of competing in foundational模型 has become prohibitively high for most startups. The paths of the original "Six Tigers" have diverged: some like智谱 and MiniMax achieved high valuations via IPOs, effectively closing the capital window for new通用模型 players. Others, like 01.AI and Baichuan, are retreating from the通用模型 race to focus on商业化 and垂直场景. The deeper change is China's AI sector accepting that its comparative advantage may not lie in foundational model突破 but in applications, engineering, commercialization speed, and integrating AI into real-world industrial and user scenarios—turning AI into a viable industry. Li and Wang, veterans from the互联网 era, represent a generation that entered with理想主义 but is now pragmatically adjusting to reality. Their strategic转身 signifies a交棒 from the狂热造神 phase to a more mature stage focused on sustainable business,合同, and现金流. This isn't a story of failure, but a体面告别 to unrealistic expectations, with the long-term battle ahead passed to a new generation of AI-native builders.

marsbit05/29 01:30

Li Kaifu and Wang Xiaochuan Pivot: The First Half of the Large Model Entrepreneurship Era Ends

marsbit05/29 01:30

Token Budget Wars: Enterprise AI Enters the 'Accounting Era'

Token Budget Wars: Enterprise AI Enters the "Accounting Era" Enterprise AI is shifting from the question of "whether to adopt" to "how to account for it." As AI inference costs evolve from experimental budgets into ongoing operational expenses, CEOs and CFOs are demanding proof of value: what tangible results does each dollar spent on tokens deliver? The core of "Token Budget Wars" is not simply about reducing AI bills, but about intelligently allocating compute resources. It involves determining which business processes warrant more computational power, which tasks can use cheaper models, which can be outsourced or handled manually, and which are merely inefficient consumption. A key insight is that AI usage (token consumption) does not equal value. While SaaS usage indicated software adoption, AI token usage only indicates the "meter is running." The same workflow can cost vastly different amounts due to factors like prompt quality, context, model choice, and retries. The critical metric for scaling is "marginal token utility"—the business value created per additional dollar of inference cost. However, this is difficult to measure due to challenges like the long tail of retries, context inflation (where costs can scale quadratically with context length), and inefficient model routing (defaulting to the most powerful model for all tasks). The competition for token allocation is intensifying because, in the AI era, influence is tied to how much intelligence one can command, not just team size. AI spending is essentially competing with labor costs, whether for replacing external BPOs, internal staff, or generating new revenue. BPO contracts provide a clearer benchmark as they are priced per completed unit. The missing layer is attribution from tokens to business outcomes. Companies need a system that connects inference spending to completed work and results, capturing the agent's decision trajectory—what it saw, retrieved, tried, and why it succeeded or failed. This recorded rationale becomes a valuable asset. Ultimately, those who master token-to-outcome attribution will control the allocation of AI resources within enterprises, deciding which workflows get more compute, which are capped, or which revert to humans. The first phase of enterprise AI proved models could do the work. The next phase will determine how much of that work is worth paying for.

marsbit05/28 12:13

Token Budget Wars: Enterprise AI Enters the 'Accounting Era'

marsbit05/28 12:13

A 10,000-Word Interpretation of the "Optical Interconnect" Industry Chain: The AI Infrastructure Bottleneck Obscured by GPU Glare

**Summary: The Rise of Optical Interconnect in AI Infrastructure** This analysis explores the critical, yet often overlooked, role of optical interconnects in large-scale AI data centers. While GPUs provide raw computational power, the efficiency of AI clusters depends heavily on high-speed data transfer between thousands of cooperating GPUs during both training and inference tasks. Copper-based electrical connections are hitting physical limits in bandwidth, distance, and power consumption. Fiber optics, using light signals, offer a superior solution with exponentially higher bandwidth and lower energy use over longer distances. This shift is driving rapid growth in the optical interconnect market. The core translation device is the pluggable optical transceiver (or module), which converts electrical signals from GPUs into optical signals for fiber transmission and vice versa. Its manufacturing involves two distinct semiconductor domains: indium phosphide (InP) for optical chips (lasers, modulators, detectors) and silicon for digital signal processing (DSP) chips. A transformative next-generation technology is Co-Packaged Optics (CPO). CPO moves the optical engine (a silicon photonic integrated circuit, or PIC) much closer to the GPU or switch inside the same chip package, drastically reducing power loss and latency. CPO necessitates an external laser source and relies on silicon photonics (using Silicon-on-Insulator/SOI wafers) for integration with silicon chips. The optical interconnect ecosystem is highly fragmented, unlike the concentrated GPU market. Key bottlenecks and players span the entire supply chain: InP substrates (e.g., AXT), epitaxial wafers (e.g., IQE), laser chips (e.g., Sivers, Lumentum, Coherent), silicon photonics foundries (e.g., Tower Semiconductor), SOI wafers (e.g., Soitec), DSP/switch chips (e.g., Broadcom, Marvell), and underlying fiber (e.g., Corning). The article posits that AI infrastructure competition is extending from "who has more GPUs" to "who can secure the scarce optical interconnect supply chain." CPO represents the largest potential growth variable, with projections suggesting it could become a market worth tens of billions of dollars by 2028. Investment opportunities vary from conservative (large, diversified players) to aggressive (small, high-beta companies focused on specific bottleneck technologies), but the sector carries significant volatility and execution risks.

marsbit05/28 11:03

A 10,000-Word Interpretation of the "Optical Interconnect" Industry Chain: The AI Infrastructure Bottleneck Obscured by GPU Glare

marsbit05/28 11:03

The Wind of 'Proactive' AI Blows into Silicon Valley: Hark Secures $700 Million in Funding

Hark, an AI startup founded in late 2025, has raised $700 million in Series A funding at a $6 billion valuation. Led by Parkway Venture Capital with participation from NVIDIA, AMD Ventures, Intel Capital, Qualcomm Ventures, and Salesforce Ventures, the company aims to develop next-generation human-computer interfaces using a combination of proprietary foundational models and custom-built AI-native hardware. Founded by serial entrepreneur Brett Adcock, Hark envisions a system of multimodal devices equipped with agentic capabilities, end-to-end voice models, and personalized memory. This "active" AI approach seeks to move beyond passive chatbots, creating collaborative companions that anticipate needs and interact naturally within the real world. Adcock's experience with Figure, a humanoid robotics company, informs this hardware-focused venture. The article argues that while current AI is powerful, it remains confined to screens and traditional interfaces like chat. The next paradigm shift requires dedicated hardware that is always-on, possesses persistent memory, and enables intuitive interaction, potentially rivaling the impact of the iPhone. Hark is assembling a team with talent from Apple, Meta, Google, and Tesla to tackle this complex engineering challenge across models, hardware, and interaction design. Finally, the piece suggests Chinese startups may have an advantage in this "active" AI hardware space due to strong manufacturing ecosystems, a vast domestic market, and supportive government policies, framing the competition as one that requires integrated progress in models, operating systems, and devices.

marsbit05/28 10:22

The Wind of 'Proactive' AI Blows into Silicon Valley: Hark Secures $700 Million in Funding

marsbit05/28 10:22

Competitors Going Public, Kimi Can't Sit Still

Competitors Go Public, Kimi Feels the Pressure Yue Zhi An Mian (Moonshot AI), the company behind the AI assistant Kimi, has begun dismantling its VIE and red-chip structure, clearing a key obstacle for a potential Hong Kong IPO. This marks a significant shift from six months ago when founder Yang Zhilin stated the company was in "no hurry" to list. The move comes as rivals like Zhipu AI and MiniMax have successfully listed on the Hong Kong Stock Exchange in early 2026, experiencing massive surges in market value. This has reset valuation logic for AI companies, turning "going public" from an end goal into a competitive necessity. Analysts suggest Kimi is both seizing a favorable market window and responding to competitive pressure. Kimi's valuation has skyrocketed from around $3 billion at its 2023 founding to over $20 billion by May 2026. Capital is betting on its potential as a future AI platform and gateway, though some caution this "emotional valuation" depends on sustained technological leadership and successful commercialization. Traditionally focused on core model R&D over user growth, Kimi has recently pivoted strategy. While its monthly active users declined through 2025, it shifted focus to Agent development and reducing marketing spend. The release of its K2.5 model in early 2026 reportedly generated substantial revenue, with annual recurring revenue reaching $200 million by April, driven by subscriptions and API services. A $2 billion D-round financing in May signaled investor approval of this commercial shift. However, listing will bring new pressures. Experts predict a listed Kimi would face stricter scrutiny on financial controls, compliance, and R&D efficiency. The narrative must evolve from pure technological breakthroughs to demonstrating clear commercialization paths, sustainable income, and a defensible valuation, balancing model superiority with business performance.

marsbit05/28 10:02

Competitors Going Public, Kimi Can't Sit Still

marsbit05/28 10:02

活动图片