LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-10-21154 posts

Atualizar

DeepSeek Open-Sources DeepInfer, Its Production Inference Framework Supporting Nvidia Blackwell and Huawei Ascend

Tópico · DeepSeek推理框架Material
2026-10-21 07:11 GMT+8

Chinese AI startup DeepSeek on October 20, 2026, open-sourced DeepInfer, the inference framework powering its own production services. The framework aims to solve the challenge of serving large models to a massive user base at a reasonable cost. According to DeepSeek, one cluster currently serves 155 trillion tokens per day, generating approximately $38 in revenue per GPU per hour at off-peak pricing.

For now, DeepInfer supports only a single model, DeepSeek-V4.1-Flash, defaulting to the updated 1017 version released over the weekend. The framework is compatible with Nvidia Corp.'s SM100 Blackwell GPUs and Huawei Technologies Co. Ltd.'s Ascend 950 neural processing units, but does not support Hopper or other older Nvidia architectures. Two features ensuring identical outputs across runs work only on Nvidia hardware. All published performance benchmarks were run on Nvidia chips, and the figures are self-reported by DeepSeek.

Technically, DeepInfer employs a speculative decoding system called DSpark, where a draft model proposes five tokens at a time and a confidence head predicts which ones the main model will accept. This reduces the number of draft tokens checked per step by about 30%, yielding an estimated 4% to 7% throughput gain at production load. A second technique claims to nearly halve total prefill computation by running only the last 128 tokens of each prompt back through the model's decoder layers.

DeepSeek has not yet released its key-value cache management and load-balancing code, stating these are tightly coupled to internal infrastructure. Splitting prefill and decode across separate machines is also left to operators. DeepInfer is available on GitHub under the Apache 2.0 license, accompanied by a deepseek-recipe library that converts requests into token sequences using the same logic as DeepSeek's official API.

Fontes:siliconangle.com

Pesquisa

Modal Launches Sticky Sessions to Solve State Continuity for Real-Time AI Apps

Material
2026-10-20 00:00 GMT+8

When building real-time AI applications such as speech-to-text models, conversational agents, or multiplayer games, developers face a core contradiction: traditional serverless computing assumes each request is independent to allow flexible scaling, but these applications require state continuity throughout a session, meaning users must return to the same processing container upon reconnection.

To address this pain point, cloud infrastructure company Modal launched the Sticky Sessions feature on October 19, 2026. Built on its low-latency Modal Servers, this feature ensures that all requests within the same session are routed to the same container. Developers can enable it by creating a server with the @modal.sessioned() decorator and starting sessions via SDK or HTTP API.

The core innovation of Sticky Sessions lies in using 'sessions' rather than 'requests' as the unit of concurrency for auto-scaling. This means the system automatically provisions new containers as active sessions increase and recycles them when sessions become idle. This mechanism avoids the previous anti-pattern of setting max_containers=1 to force stickiness at the cost of scalability, and also solves the issue where WebSocket disconnections could route to a new container, causing state loss.

For teams developing low-latency voice agents, shared chat rooms, or online games, this update provides first-class infrastructure support. It allows applications to maintain stable WebRTC signaling and continuous memory state while still benefiting from the pay-as-you-go and elastic scaling advantages of serverless architecture.

Fontes:aihot.news

Pesquisa

Musk says Tesla and SpaceX form independent committees to explore merger, requiring 50% shareholder vote

Tópico · 特斯拉SpaceX合并Estrutural
2026-10-21 06:16 GMT+8

Elon Musk stated during SpaceX's Tuesday earnings call that both Tesla and SpaceX have created special committees of independent directors to discuss a potential merger. This represents the most detailed update investors have received regarding the idea of a tie-up, which has been the subject of much chatter since SpaceX went public earlier this year.

Musk, who helms both companies and owns billions of dollars in stock of each, said he cannot be part of those exploratory groups. "Do I think it would be a good idea? Yes," he said on the call. "But at the end of the day, the shareholders have to be supportive of this, and particularly on the Tesla side, we will need a 50% vote of all shares, not just those who vote, but of all shareholders." He described clearing that bar as "quite a high hurdle."

While Musk sees clear benefits in combining the companies, citing their collaboration on AI software, chip design, manufacturing, robotics, and Starlink connectivity, he cautioned investors against assuming the outcome. "I don't think we should necessarily assume that a SpaceX and Tesla merger is a certainty," he added. "That would be overstating it." If the companies ultimately decide to merge, it would further consolidate Musk's business empire, often referred to as 'Musk Inc.' Previous consolidations within the empire include xAI's acquisition of X in 2025 and SpaceX's acquisition of xAI in February.

Following the news, SpaceX stocks rose approximately 6.6% and Tesla rose 1.6% in after-hours trading.

Fontes:businessinsider.com

Pesquisa

Anthropic H1 revenue reaches 15x prior-year level at $16.3B; net loss $80.7B

Tópico · Claude企业采用率Material
2026-10-21 06:48 GMT+8

ETNews reports, citing The Wall Street Journal’s account of an S-1 document it obtained, on Anthropic’s first-half financial results. This article relies on the ETNews report, without direct access to that document, and does not establish that it was filed with the U.S. Securities and Exchange Commission (SEC). The reported revenue reached $16.3 billion—15 times the same period last year—while the reported net loss was $80.7 billion.

According to the relay report, approximately $72 billion of this massive loss was not due to daily operating expenses but rather book losses resulting from value fluctuations in financial instruments such as convertible bonds issued for fundraising. This describes a non-cash component of the reported loss; it does not establish the company’s actual cash flow conditions.

Underpinning Anthropic's rapid growth are substantial infrastructure investments. The relay report states that spending on compute capacity and infrastructure reached $9 billion in the first half alone, exceeding the $7.3 billion spent in all of last year. These are cumulative spending figures for a half year and a full year, respectively, and cannot be used to infer year-on-year growth rates or changes in cash flow. It also reports that Anthropic has signed irrevocable future compute procurement contracts totaling over $450 billion, primarily with Google Cloud, Amazon Web Services, and Microsoft. Including chip leasing agreements via Broadcom and contracts with smaller cloud providers, total future infrastructure commitments could exceed $600 billion.

The report also states that Anthropic noted high customer concentration, with about one-fifth of its revenue coming from just two clients. The company also highlighted technical risks, including the potential for AI models to exhibit 'self-preservation' behaviors or generate coercive responses. These relayed figures describe Anthropic's revenue, loss composition and future compute-contract obligations; they do not establish actual cash flow, an IPO valuation or the profitability outlook for the whole sector.

Fontes:etnews.com

Pesquisa

Janus Henderson Credit Head John Lloyd Sees CoreWeave Bond Sell-Off as Overreaction

Tópico · CoreWeave动态Material
2026-10-21 06:03 GMT+8

John Lloyd, global head of multi-sector and corporate credit at Janus Henderson, believes the recent sell-off in CoreWeave bonds is an overreaction. Speaking to CNBC, he noted that bond prices have fallen, spreads have widened, and yields have risen, but he sees no fundamental change in the ecosystem that would warrant such a move, other than higher financing costs—which he believes CoreWeave can address creatively.

Lloyd's Janus Henderson Multi-Sector Income Fund holds CoreWeave's 9.75% bonds due in 2031, now yielding 15.97% to maturity, a more than 70% jump since issuance in April. CoreWeave's five-year credit default swap spread stands at 1,085 basis points, with levels above 1,000 generally signaling credit distress; its shares have fallen roughly 20% over the past week. Total debt is approximately $51.6 billion, including about $35 billion in notes, bonds and loans. The debt-to-equity ratio has risen from 290% a year ago to 702%. According to consensus estimates, CoreWeave will spend about $36.4 billion on capital expenditures this year against EBITDA of $7.5 billion.

Lloyd's argument does not deny the severity of these figures; rather, it explains why he believes the market price has overshot what fundamentals warrant. He outlines three specific mechanisms: first, CoreWeave can borrow against its GPU assets or rental contracts without relying entirely on the high-yield market; second, the company can ask its largest customers—including Microsoft, Meta Platforms and OpenAI—to increase prepayments; third, CoreWeave has no near-term maturities, meaning it faces no refinancing pressure in the immediate future. He said:

'They can probably get creative enough where they don't even need to come to the high-yield market.'

On the demand side, Lloyd sees the AI compute supply-demand dynamic remaining tight. He noted that adoption rates within the ecosystem are running above consensus demand curves, while chips and memory face supply bottlenecks—a favorable setup for CoreWeave. He also pointed out that if Anthropic proceeds with its expected November IPO, followed by OpenAI, successful listings by frontier labs could ease market concerns about AI demand and indirectly support neocloud operators' credit pricing.

It is worth noting that Lloyd himself holds these bonds, so his judgment carries a conflict of interest. The article also notes that CoreWeave still needs to convince markets it can handle higher rates and get through financing bottlenecks; Lloyd sees the company making it, but acknowledges the bond recovery will not be a straight line. If AI compute demand growth slows, or if major customers reduce their willingness to prepay due to their own budget cycles, the multiple funding channels Lloyd relies on could tighten simultaneously—and the 702% debt-to-equity ratio means any further rise in financing costs will directly erode shareholder returns.

Fontes:cnbc.com

Pesquisa

Amazon Pledges Over €16 Billion Investment in Germany for 2027, Focusing on AWS Data Centers and AI Infrastructure

Tópico · 亚马逊德国投资计划Material
2026-01-14 01:17 GMT+8

US tech giant Amazon announced on October 20, 2026, at the "Invest in Germany Summit" in Berlin that it plans to invest over €16 billion ($17.9 billion) in Germany in 2027. The funds will primarily support the expansion of Amazon Web Services (AWS) data centers and artificial intelligence infrastructure, while also covering its Leo satellite internet service, logistics upgrades, and new Prime Video content production.

According to David Zapolsky, Amazon's Chief Global Affairs and Legal Officer, this latest move builds on more than €100 billion invested in Germany since 2010. Approximately €4.8 billion is earmarked specifically for expanding AWS cloud and AI infrastructure in the Frankfurt area to meet growing demand for AI services. Additionally, Amazon is opening a second robotics and AI lab in Berlin.

German Federal Minister for Economic Affairs and Energy Katherina Reiche welcomed the announcement, stating that Amazon's investment is a clear vote of confidence in Germany as a location for the future, helping to strengthen the country's position in international technology competition and create jobs. The announcement comes amid efforts by the German government to attract high-tech investments to revive its stagnating economy.

Fontes:aboutamazon.com

Pesquisa

Shu Junliang of Qwen Office: The Battleground for Enterprise AI is Customer Choice; Context Should Be an Independent Infrastructure

Tópico · 千问办公Material
2026-09-27 16:38 GMT+8

During the September 2026 Yunqi Conference, Shu Junliang, head of "Qwen Office"—Alibaba's newly consolidated AI office brand—gave a group interview addressing the fierce competition among global tech giants in enterprise office scenarios. Shu presented several core judgments regarding the implementation of enterprise-level AI. He argued that although underlying technologies and user needs are changing rapidly, the true battleground at this stage is not product volume, but the actual choice and depth of use by enterprise customers.

Regarding Alibaba's strategic move to integrate teams like QoderWork and DingTalk Wukong into Qwen Office, Shu confirmed that internal efforts are converging toward a unified capability outlet. However, he emphasized that Qwen Office should not be defined as a fixed desktop agent or entry point, but as a capability system. In relation to traditional office software like DingTalk, the two are not simply replacing each other but undergoing bidirectional integration: DingTalk's accumulated data assets and business logic will serve as the foundation, while Qwen Office provides stronger natural language interaction and generalization capabilities. Whether they eventually merge into a single product will be determined by user demand and industry evolution.

On the technical architecture level, Shu proposed a key viewpoint: "Context should follow the subject, not a specific Agent product." He believes that enterprise context should become a data infrastructure independent of specific applications, belonging to the enterprise itself. This means that whether an enterprise chooses to develop its own AI or procure other products in the future, it should be able to derive capabilities from this unified context. This decoupled design aims to solve the problem of data silos between different AI tools, allowing context to become a universal digital asset for enterprises, much like a database.

Additionally, Shu analyzed the essential differences between enterprise-level Agents and personal assistants. He pointed out that personal assistants typically represent the user themselves, with operations under individual control. Once an Agent enters organizational processes, it must answer questions such as "Who brought it in," "What can it access," and "Who is responsible if something goes wrong." Therefore, identity, permissions, data ownership, auditing, and responsibility constitute the basic requirements for enterprise AI. In terms of product implementation, Qwen Office emphasizes pre-defined boundaries, execution within sandboxes, manual confirmation for high-risk operations, and post-auditing to meet enterprise security and compliance needs.

Although full-stack AI capabilities are seen as Alibaba's phased advantage, Shu remained cautious, acknowledging that competition has just begun and no fundamental differentiation exists yet. He admitted that while pay-for-results is the ideal direction, it is difficult to standardize measurement in enterprise scenarios. Current pricing models are more a result of the industrial stage; as software homogenization alleviates, the scarcity customers are willing to pay for will shift from pure model costs to truly differentiated software design and functionality.

Fontes:geekpark.net

Pesquisa

Goldman sees Big Tech spending $1.2 trillion on AI infrastructure in 2027

Tópico · 五大科技巨头AI资本开支Material
2026-09-27 16:17 GMT+8

According to Bloomberg, Goldman Sachs strategist Ryan Hammond projects that Amazon, Alphabet, Microsoft, Oracle and Meta will spend a combined $1.2 trillion on AI infrastructure in 2027.

That is more than 50 percent above the roughly $800 billion projected for this year, and above Wall Street's consensus of $1.1 trillion. Goldman says that measured against GDP, it would be the biggest investment cycle since railroad construction in the 19th century.

Growth is slowing, though: from nearly 100 percent in 2026 to 54 percent in 2027 and 12 percent in 2028. Goldman also notes that spending now exceeds what the companies generate from operations, meaning more debt financing; recouping the outlays would require about $300 billion a year in AI revenue, and bottlenecks in power, labor and memory chips could slow the pace further.

Fontes:the-decoder.com

Pesquisa

Why DeepSeek's Liang Wenfeng Rarely Signs Flagship Model Reports, Only Foundational Technique Papers

Tópico · 梁文锋底层技术路线Material
2026-09-25 18:48 GMT+8

Over the past three years, DeepSeek has released multiple flagship models, each technical report carrying over a hundred author names. But look at the author lists and you find an interesting pattern: Liang Wenfeng's name rarely appears on V3.2, V4, or V4.1-Flash reports. Instead, he consistently appears on a different class of papers—MLA attention mechanism, MoE routing strategy, mHC network topology, DSpark inference scheduling, and DSec sandbox infrastructure. These are the atomic technical units that constitute model capability, not the models themselves.

Lei Feng Net author Gao Yunyi compiled this observation into an article, calling it a technical sniping route, and counted the 11 foundational papers Liang Wenfeng signed. The signing pattern itself is verifiable: check the author lists on arXiv. What it reveals is that Liang Wenfeng's personal focus is not training a larger model but finding the deepest bottleneck that makes a model cheaper or stronger.

Take MLA (Multi-head Latent Attention) as an example. When large models process long text, they must cache massive key-value pairs (KV Cache), which directly determines memory usage and speed during inference. MLA's approach is to compress this cache with low-rank vectors, reducing KV Cache volume by approximately 93.3%. This is not optimization at the model level but a redesign of the data structure for one specific component within the attention mechanism. Similarly, MoE routing solves the scheduling problem of which experts each token should invoke; NSA solves the memory-access problem of how to align sparse computation with GPU hardware circuits; mHC solves the topological constraint problem of whether numerical stability will collapse during very deep network training.

These papers share a common characteristic: they attack foundational bottlenecks that the industry acknowledges but has not effectively solved, and their solutions land at the architecture or mathematics level, not the data-volume or compute-investment level. Gao Yunyi cites a comparison in the article: OpenAI's GPT-4-class model training costs hundreds of millions of dollars, while the DeepSeek-V3 team reported completing training on 2,048 H800 GPUs for approximately $5.6 million. The article also relays SemiAnalysis's caveat—that the $5.6 million is merely the direct electricity bill for pure GPU training, excluding R&D losses, hardware depreciation, and total infrastructure investment. But even with the corrected scope, the order-of-magnitude difference still holds, and its source is precisely the foundational architectural choices described above.

However, what the signing pattern can prove requires caution. Liang Wenfeng's choice to sign foundational papers rather than model reports may reflect personal research preference, or it may be a natural distribution of authorship contribution—foundational technique papers have fewer authors, model reports have more. Gao Yunyi interprets it as a conscious sniping strategy, which is a reasonable inference but not a proof. What actually supports this inference is that these foundational choices repeatedly hit industry bottlenecks and produced verifiable efficiency gains, not the signing behavior itself.

Fontes:leiphone.com

Pesquisa

U.S. Office of Financial Research: Updated Estimate Suggests Treasury Repo Clearing Rule Could Raise SOFR by About 1 Basis Point

Tópico · SEC回购集中清算规则Material
2026-10-20 12:00 GMT+8

The U.S. Office of Financial Research (OFR) published an updated estimate in October. Under the assumption that the SEC's rule requiring certain Treasury repurchase agreements to enter central clearing does not change participant behavior, the clearing mandate could add slightly more than $500 billion in daily transaction volume to SOFR (the Secured Overnight Financing Rate), while the hypothetical SOFR could rise by only about 1 basis point above the published rate, with volatility roughly unchanged.

SOFR is the benchmark rate for U.S. dollar short-term funding, calculated daily by the Federal Reserve Bank of New York based on actual Treasury repo transactions. The SEC's rule requires certain bilateral Treasury repos to be centrally cleared through FICC's (Fixed Income Clearing Corporation) delivery-versus-payment (DVP) service, with a compliance deadline of June 30, 2027. This means non-centrally cleared bilateral repos (NCCBR) that were previously outside SOFR's calculation would enter the FICC DVP segment and be included in SOFR's transaction volume pool.

OFR's method applies the SEC rule and the Federal Reserve Bank of New York's SOFR eligibility criteria to NCCBR data, assuming all repo volume subject to the clearing requirement is centrally cleared through FICC DVP, and assuming no change in market participant behavior. Compared with the April 2025 preliminary estimate based on three days of pilot data, this update uses a full year of NCCBR data, raising the estimated new volume from about $85 billion to slightly more than $500 billion. Despite this large increase in volume, the hypothetical SOFR could rise by only about 1 basis point above the published rate, with volatility roughly unchanged. This conclusion holds under the premise that details of the SOFR calculation remain unchanged.

Beyond the volume change from central clearing itself, OFR identified a second channel: clients already within the centrally cleared system could migrate between clearing segments. Money market funds, the primary cash lenders in repo, have historically cleared through Sponsored DVP but have begun using Sponsored General Collateral (GC). Under an extreme scenario in which all these funds' volume shifted from Sponsored DVP to Sponsored GC—which is not included in SOFR's calculation—SOFR could see a relative shift of up to 2 basis points above historical levels.

OFR explicitly states that this is an approximate backward-looking analysis, not a forecast of the market after the transition occurs. The core assumption is no change in participant behavior, but in reality balance-sheet-constrained dealers may reprice, while netting in central clearing could ease those constraints. The full effects remain uncertain and depend on how participants choose to clear. OFR will continue to monitor these effects as the compliance date approaches.

Fontes:financialresearch.gov

Pesquisa

AI Agent Cloud Provider Daytona Raises $60M Series B at $1B Valuation

Tópico · Daytona融资动态Material
2026-10-20 20:00 GMT+8

San Francisco-based startup Daytona announced the completion of a $60 million Series B funding round led by FirstMark Capital, bringing its valuation to $1 billion. This round increases Daytona's total funding to approximately $91 million. Investors also include founders from Stripe, Shopify, and Perplexity, as well as strategic investors like Qualcomm Ventures, Samsung Next, and Okta Ventures.

Daytona builds an "agentic cloud," providing dedicated virtual computer environments (sandboxes) for AI agents. Unlike traditional cloud services, Daytona claims it can configure a sandbox in 60 milliseconds, whereas providers like AWS or Azure typically take between 30 seconds and several minutes. This low latency is critical for applications requiring thousands of parallel agents.

According to Fortune, Daytona co-founder and CEO Ivan Burazin stated that the company's revenue run rate grew 25-fold between January and September 2026, averaging 50.9% monthly growth. Leading AI companies including Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta, and Cohere currently use Daytona's services for model training and testing.

Burazin distinguishes Daytona from "neocloud" providers like CoreWeave: the latter focus on GPUs for training and running AI models, while Daytona focuses on CPUs for supporting agent execution. He predicts that spending on CPUs will eventually exceed spending on GPUs due to the explosion of agentic applications, potentially leading to a severe CPU shortage. This view aligns with previous statements by AMD CEO Lisa Su regarding increased server CPU demand driven by agentic AI.

Fontes:fortune.com

Pesquisa

Meta Releases Open-Weight Model Muse Spark Open, Focusing on Long-Context Coding and Agentic Work

Material
2026-10-15 08:00 GMT+8

Meta released the open weights for Muse Spark Open on October 15, 2026, announcing its immediate availability on platforms like Fireworks AI. The model is a Mixture-of-Experts (MoE) architecture with 816 billion total parameters and 42 billion active per token, featuring a 1 million token context window designed for long-horizon coding and coherent agentic workflows.

According to benchmark results published by Fireworks AI, Muse Spark Open demonstrates strong performance in specialized domains such as law and finance. On Vals AI’s Harvey Legal Agent Benchmark leaderboard, it ranked first among 75 models with a pass rate of 25.4%, nearly four times that of Claude Opus 5 (6.7%). On Rogo’s Big Finance Bench, it outperformed Claude Opus 5.5 at every difficulty level and narrowly edged out GPT-6 Astra on the hardest tasks. Additionally, its scores in clinical documentation tasks were comparable to leading closed-source models but at lower token prices.

Beyond cloud services, Meta highlighted the model's local deployment capabilities. In collaboration with Apple, Muse Spark Open was optimized to run on two Mac Studios equipped with M5 Ultra chips, leveraging unified memory architecture to support large model inference. This flexibility allows enterprises to deploy high-performance AI in private environments while meeting data residency and security compliance requirements.

While the vendor claims this is one of the most powerful US open-weight models to date, some performance data comes from vendor self-tests or specific partner evaluations. Developers can download the weights via Hugging Face or use them on cloud platforms such as Amazon SageMaker AI and Google Cloud Model Garden.

Fontes:research.meta.ai

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

For readers, the practical value of this observation is that it reduces DeepSeek's cost-performance label to a traceable technical route. If you care about AI infrastructure or model training efficiency, what deserves attention is not the parameter count of the next large model, but the next foundational bottleneck to be conquered—attention data structures, expert scheduling, network topology constraints, inference compute-memory access alignment, and agent training environment isolation. Liang Wenfeng's signing pattern at least suggests: someone has been consistently working in these directions.