ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-2082 posts

Refresh

Intel Reclaims Overall x86 CPU Market Share in Q3 2026 Driven by Desktop Gains; AMD Sets Records in Mobile and Server Segments

Material
2026-11-20 06:10 GMT+8

According to Mercury Research data reported by TechRadar, Intel reclaimed overall x86 CPU market share from AMD in Q3 2026 for the first time since Q2 2025. Intel's share of x86 processor unit shipments (excluding embedded and semi-custom chips) rose to 69.9%, while AMD's declined to 30.1%. This represents a sequential shift where AMD lost approximately 0.6 percentage points.

The gain was concentrated almost entirely in the desktop segment, where Intel's share jumped roughly five points to 70.1%, pulling AMD down from 34.9% to 29.9%. Conversely, AMD continued to expand its presence in other key areas: its mobile share reached a record 29.2% (up from 28.9%), and its server share hit a record 35.5% (up one full point). When narrowing the server comparison to EPYC against Intel's Xeon SP line, AMD's unit share reaches an impressive 46.8%.

Despite Intel's desktop surge, both companies face headwinds. Intel expects client revenue to fall by more than 20% sequentially in Q4, and AMD guides its combined Client and Gaming segment down by more than 20% as well, largely due to memory price pressures squeezing PC costs. Furthermore, AMD has guided Q4 revenue to around $17 billion (±3%), while Intel expects $15 billion to $16 billion, meaning AMD is projected to out-earn Intel for the quarter regardless of who ships more desktop chips.

Sources:TechRadar

Research

Fireworks AI Co-founder Chenyu Zhao Proposes a Two-Model World: Companies Spending Over $10,000 Monthly on Tokens Should Consider Switching to Open-Source Models

Material
2026-11-20 08:26 GMT+8

Chenyu Zhao, co-founder of Fireworks AI, presented a concrete financial threshold for enterprise AI expenditure at the Slush startup conference in Helsinki, Finland: when user scale exceeds 10,000 to 100,000 and monthly token spending reaches $10,000 to $100,000, it is time to seriously consider switching from frontier models to open-source models. His core argument is not that open-source models have caught up to frontier models in capability, but that most enterprise workloads do not require the general intelligence level provided by frontier models.

Zhao described the future landscape as a 'two-model world': frontier models are used for high-level planning and complex reasoning, while open-source models handle daily execution, operating in parallel. The key mechanism connecting them is routing—assigning tasks to the most suitable model based on cost rather than per-token unit price. He stated that the industry is no longer focused on the cost per token, but on the total cost to complete a single task. This distinction has practical implications for enterprises: an open-source model may have a lower unit price, but if it requires more tokens to complete a task, the total cost might not be lower; conversely, a frontier model with a higher unit price might be cheaper overall if it completes the task in one go.

Zhao’s premise is that the capability gap between open-source and frontier models is currently at most six months, and likely closer to three. He noted that a wave of excellent open-source models emerged following DeepSeek, significantly narrowing the gap from what was once one to one-and-a-half years. However, this assessment comes from his own evaluation and is not independently measured. Notably, Anthropic released Opus 5.5 in September, which reduced costs for typical workloads by 40%, lowering input and output token prices to $4 and $20 per million, respectively, and cache reads to $0.20 per million. As frontier models themselves continue to drop in price, Zhao’s suggested switching threshold may shift upward over time.

Zhao’s framework carries a conflict of interest that readers must evaluate: Fireworks’ business is precisely helping enterprises fine-tune open-source models on their own data and host them. Its customers include Uber, Shopify, and Notion. In autumn, Fireworks disclosed two security incidents where hackers accessed customer data through its cloud provider; Notion subsequently disconnected from the service, and Uber reported that personal information for approximately 150,000 drivers was stolen. Zhao acknowledged these incidents and stated that the company is migrating batch prediction services to online processes to reduce data retention times. While these accidents do not affect the logic of his cost framework, they impact whether enterprises are willing to entrust their data to intermediary platforms like Fireworks.

In summary, Zhao’s proposed routing framework for a two-model world is an actionable decision-making tool: first evaluate based on task cost rather than per-token unit price, then decide whether to introduce open-source models based on expenditure scale. However, its premise—that the open-source model gap is only three to six months—is Zhao’s judgment rather than independent verification, and continued price cuts by frontier models may alter the economics of switching. Enterprises adopting this framework need to measure task-level costs themselves rather than directly applying his thresholds.

Sources:TechCrunch

Research

Luxury Resale Platform The RealReal Expands AI Shopping Agent and Deploys Backend Automation

Quick take
Verified 2026-11-20 09:08 GMT+8

US-based luxury resale platform The RealReal (Nasdaq: REAL) announced on October 8, 2026, that it is expanding 'Ask TRR,' an AI shopping agent built with Google Cloud's Gemini Enterprise, to all 45 million members. The tool addresses the search challenges inherent in 'one-of-one' inventory by allowing consumers to describe needs in natural language—such as occasion, style, or era—rather than relying on precise brand or designer keywords.

On the supply side, The RealReal has introduced 'Athena,' a backend automation system. Athena uses image data and proprietary information to generate product descriptions, provide pricing suggestions, and assess counterfeit risks. According to TechOrange, the system is expected to handle nearly half of all items by the end of this year, up from 35% at the end of 2025. This triage mechanism allows low-risk items to bypass some manual authentication steps, while high-risk goods still undergo final verification by professional authenticators using X-ray equipment and tactile inspection. Company executives expect this system to reduce the average time from warehouse intake to live listing from 14 days to 7 days.

CEO Rati Sahi Levesque stated that the platform offers over 1.5 million unique pieces daily, backed by 15 years of proprietary data. The core value of Ask TRR lies in translating vague consumer intent into precise product matches, thereby accelerating turnover and reducing markdown losses from stagnant inventory. While the company claims the pilot was successful, it has not yet published specific conversion rate or sales growth data.

Sources:Google Cloud Press Corner

Research

Samsung Electronics Reshuffles Semiconductor Leadership: Three VPs Promoted to President as Top Executives Shift to Future Businesses

Material
2026-11-20 08:16 GMT+8

South Korean tech giant Samsung Electronics announced its "2027 Regular President-Level Personnel Changes" on November 20, 2026, involving six executive positions. The core of this adjustment lies in the Device Solutions (DS) division, where the company promoted three Vice Presidents who led record performance and development competitiveness to President, aiming to boost organizational dynamism through generational transition.

Hwang Sang-jun was promoted from Head of Memory Development and DRAM Development Lab to President of the Memory Business Unit, retaining his role as Head of the DRAM Development Lab. He previously led efforts to secure development competitiveness across all memory products, including DRAM and High Bandwidth Memory (HBM). The company expects him to drive the paradigm shift in the memory business for the AI era.

Cho Sang-yeon was promoted from Head of DSA (Semiconductor Americas Corporation) to President of the System LSI Business Unit. Since December 2024, he has overseen Americas operations and helped drive the DS division's historic best performance this year. His new role focuses on improving the System LSI structure and securing new customers.

Lee Jong-myung was promoted from Head of Process Development at the Semiconductor Research Institute to President and CTO of the DS Division, concurrently serving as Director of the Semiconductor Research Institute. With over 28 years of process expertise, he has led new process development for all semiconductor products since June 2022, including DRAM, NAND, and Foundry. He is expected to overcome technical limits such as miniaturization.

Concurrently, Jeon Young-hyun, Vice Chairman and Head of the DS Division, stepped down as Head of the Memory Business Unit, handing it to the new President Hwang Sang-jun, to focus on strengthening future businesses like Foundry. No Tae-mun, President and Head of the DX Division, also stepped down as Head of the Mobile eXperience (MX) Business Unit, retaining his role as Head of the RX Business Promotion Office to concentrate on preparing new growth engines like robotics.

This arrangement indicates that Samsung is delegating day-to-day core operations to verified business unit leaders while allowing top executives to focus on securing long-term growth drivers. Additionally, Choi Won-jun was appointed as the new President of the MX Business Unit. As COO and Head of Development, he successfully launched the Galaxy S26 and Z Fold 8.

Sources:미래를 보는 창 - 전자신문

Research

Candid Health's US Medical Billing Platform: What a 59% Resolution Rate Means When AI Voice Agents Must Diagnose and Prescribe in a Single Call

Material
2026-11-19 20:00 GMT+8

Candid Health, a medical billing platform in the United States, has deployed AI voice agents within its revenue cycle management workflow to call insurance companies and investigate the reasons for claim denials. In its initial deployment, these agents handled 12,860 calls, resolving 59% of them. At first glance, the fact that more than 40% of calls failed to resolve the issue might not seem impressive. However, Doug Proctor, Candid Health’s Chief Operating Officer, explains why they consider 59% a meaningful result: the key lies in the definition of 'resolution.'

Candid Health stipulates that a call is only considered 'resolved' if it accomplishes two things simultaneously: uncovering the reason for the denial and identifying the next step to fix it. Proctor stated that they deliberately set a high bar because 'getting part of the answer doesn’t really help our billers.' Billers need complete action information—knowing why a claim was denied and knowing what to do next—to free up their time for more complex work. If an agent only identifies the cause but fails to determine the solution, the biller still needs to make another call themselves, resulting in limited time savings.

This definition directly alters the meaning of the numbers. Under a loose definition—for example, counting a call as 'processed' merely if the agent connected and obtained partial information—the same system might report figures of 80% or higher. Under Candid Health’s strict definition, 59% means that in over half of the calls, the AI agent independently completed the full chain of information gathering from diagnosis to solution, without requiring biller intervention. Proctor noted that 'getting both of these from a single insurance company call is difficult even for an experienced biller.'

The transferable value of this case lies in the fact that whenever any AI voice agent vendor or client reports resolution rates, completion rates, or automation rates, the first question to ask is not about the number itself, but the task definition behind it. Does 'resolution' mean connecting the call, obtaining partial information, or completing an end-to-end workflow? Percentages under different definitions cannot be directly compared. Candid Health’s example demonstrates a conscious choice: preferring to report a lower number to ensure that the figure corresponds to a complete, operationally valuable result.

It is important to note that the 59% figure comes from Candid Health’s self-reporting, published by ElevenLabs as a customer case study. The article does not provide independent audit or scoring rule verification, so readers cannot confirm whether every call was strictly judged according to the described dual criteria. Furthermore, Candid Health’s agents are designed to investigate denials rather than appeal them, which means the scope of 'resolution' has boundaries—it does not include the initiation or execution of subsequent appeals processes.

Candid Health also revealed a notable side effect: some investigation calls uncovered upstream issues. During one investigation, an agent discovered that a healthcare provider was not credentialed with the payer on the date of service. Fixing such upstream problems can prevent tens, if not hundreds, of future denials for that customer. This indicates that voice agents operating under strict completion standards offer value beyond saving call time; they also feed structural problems discovered during investigations back into the upstream process.

Sources:ElevenLabs

Research

Dutch Enterprise Search Firm Elastic Sees Stock Slide Over 16% Despite Earnings Beat Due to Weak Guidance

Material
2026-11-20 07:33 GMT+8

Netherlands-based enterprise search software firm Elastic N.V. reported its second-quarter fiscal 2027 financial results on November 19, 2026. Although the company posted revenue and adjusted earnings per share that beat analyst expectations, its forward guidance came in below forecasts, causing the stock to decline 8% during regular trading and another 8% in after-hours sessions, for a total drop of over 16%.

The data shows that Elastic's revenue rose 17% to $496 million, surpassing the Street's target of $487 million. Earnings before certain costs such as stock compensation were 94 cents per share, easily beating the consensus estimate of 81 cents. Net income rebounded from a loss of $51.3 million a year ago to a profit of $171 million. However, guidance for the current quarter projects sales between $507.5 million and $508.5 million and EPS between 90 and 92 cents, both below market expectations of $517.6 million and 95 cents respectively. Additionally, a weak Euro is expected to cause a $4.5 million headwind on revenue this quarter.

CEO Ash Kulkarni stated that new AI products like Elastic Context Engine and AlertZero are positioned to drive significant long-term revenue growth, citing expanding customer adoption. However, investors appeared more focused on macro currency impacts and broader concerns about AI potentially challenging traditional software business models. Elastic's stock had rallied more than 150% between April and October, but recent declines have erased nearly all those gains, leaving the shares up just over 1% year-to-date.

Sources:SiliconANGLE

Research

NVIDIA Dynamo Explains Three Constraints on Reinforcement Learning Inference Bottlenecks and Corresponding Service-Layer Mechanisms

Material
2026-11-20 04:28 GMT+8

When scaled, the core challenge for the inference service layer in reinforcement learning (RL) training is not a single bottleneck, but the simultaneous interaction of three independent constraints. An NVIDIA technical blog names these three constraints rollout completion time, repeated context, and weight freshness, noting that which one to solve depends on which constraint becomes the limiting factor in the current workload.

The first constraint stems from unevenness in rollout completion times. A reasoning rollout may contain only a single model response, whereas a coding agent’s rollout might involve repository exploration, code editing, tool calls, and test execution. When a training step requires a set of samples to be fully ready before proceeding, the slowest rollout dictates the wait time for the entire step. NVIDIA points out that evenly distributing request counts across GPUs does not equate to evenly distributing actual workloads.

The second constraint is redundant computation. Reinforcement learning frequently generates multiple solutions for the same task, with these solutions sharing identical instructions, task descriptions, or code repository contexts. If every request is sent to a worker that has not cached that prefix, the GPU must repeatedly process the same input before it can begin generating new content.

The third constraint is weight freshness. During training, model weights are continuously updated. Workers must receive new weights within the sampling timing and acceptable staleness limits prescribed by the algorithm; otherwise, the generated samples cannot be used for the next training step.

As a service layer, NVIDIA Dynamo provides specific mechanisms for each of these three constraints. For the problem of redundant computation, Dynamo’s router balances KV cache reuse against worker load: it prioritizes sending requests with shared prefixes to workers that have already cached them to save compute, but it avoids dumping all requests onto a single worker, which would cause queuing delays.

For the weight distribution problem, Dynamo works with the ModelExpress component to adopt incremental transfers. Specifically, instead of sending a full model checkpoint to every worker, it transmits compressed changes (deltas). Workers reconstruct the new weights from a matching base version. Compatible workers can copy weights directly from another worker’s GPU memory via the NIXL protocol, bypassing remote storage.

For fault recovery, after detecting a worker loss, Dynamo routes new requests to healthy capacity and adds replacement workers to the routing pool once they are ready. Replacement workers can catch up with the training progress by loading a checkpoint and then sequentially applying subsequent deltas, while healthy workers continue generating rollouts during this period.

The article summarizes several customer adoption cases: Microsoft AI in the United States uses Dynamo to support rollout inference during post-training phases; US cloud provider CoreWeave states that Dynamo combined with ModelExpress accelerates model weight updates by up to 15x; US programming agent company Cognition uses Dynamo to manage inference workers and route rollouts during the post-training of SWE-1.7; Prime Intellect uses Dynamo for distributed inference orchestration; and H Company claims that Dynamo combined with ModelExpress reduces average training-step time by up to 20% and GPU-hours by up to 5%. These performance figures are self-reported by the vendors or customers, and the article does not provide independent reproduction methods.

Sources:NVIDIA Technical Blog

Research

US Senate Majority Leader John Thune Forecasts Debt Ceiling Delay to Late 2027, Forcing Trump-Democrat Negotiations

Material
2026-11-20 07:30 GMT+8

US Senate Majority Leader John Thune (R-S.D.) predicted in an interview with Semafor that the US debt ceiling issue will likely be "punted" into next year. Citing potential extra steps by the Treasury Department, Thune forecasted a final deadline in the second half of 2027.

This timeline implies that President Donald Trump will have to negotiate with a Democratic-controlled House and Senate on the debt ceiling for the first time since taking office. Thune noted that Republicans currently lack the votes to raise the debt ceiling along party lines in either chamber. While budget reconciliation could theoretically address the debt limit, Thune deemed it unlikely this year, stating that any such bill would focus on defense funding rather than avoiding default.

For investors, this forecast shifts the window for debt ceiling political maneuvering to late 2027 and defines its resolution path as dependent on bipartisan negotiation. This increases uncertainty around future fiscal policy, as divergent spending priorities between the parties may intensify during talks.

Sources:semafor.com

Research

Search and Observability Firm Elastic Announces Agreement to Acquire AI Agent Evaluation Platform HoneyHive

Material
2026-11-19 08:00 GMT+8

US-based search and observability software company Elastic announced on November 19, 2026, that it has entered into an agreement to acquire HoneyHive, an enterprise AI observability and evaluation platform. The acquisition aims to help customers build more reliable AI agents and applications by addressing the new challenges arising as AI shifts from generating responses to taking actions.

As AI agents execute decision chains with real-world consequences, traditional application monitoring often fails to detect "silent failures," such as an agent selecting the wrong tool or acting on bad information. HoneyHive's platform provides detailed tracing of complex agent workflows, alongside offline evaluation and experimentation capabilities, allowing developers to test changes before deployment and continuously monitor for unexpected behavior in production.

Elastic plans to integrate HoneyHive's agent-specific tracing and evaluation workflows into its existing Elastic Observability product. This move reflects how infrastructure vendors are actively expanding their platforms to cover critical needs in the AI agent lifecycle, following similar acquisitions by security firms like CrowdStrike. Upon closing, Elastic intends to continue supporting existing HoneyHive customers and will share more details about the consolidated product roadmap in the coming months.

Sources:Elastic Blog

Research

Google Cloud Launches Object Enrichments Preview at $700 Per 10M Items

Material
2026-11-20

US cloud provider Google Cloud announced on November 20, 2026, the introduction of a new feature called "object enrichments," currently available in preview. The service aims to address a core pain point in enterprise AI deployment: transforming petabytes of unstructured data stored in the cloud—such as images, PDFs, and videos—into "context" that AI agents can understand and reason over.

Traditionally, enterprises had to build complex custom pipelines to preprocess this data, a process that could take months and resulted in metadata being separated from the original files, increasing management overhead. "Object enrichments" uses built-in AI models (including those for image annotation and PDF classification) to automatically extract rich contextual metadata from Cloud Storage objects and write it directly into an "object context" field. This means the context lives alongside the actual bytes, traveling with the object during moves and deletions without requiring a separate metadata storage system.

Google Cloud highlights its cost efficiency, stating that processing 10 million images or PDFs costs just $700, representing a 75% saving compared to similar services from other cloud providers. This price point is significant for enterprises managing large-scale media assets or document repositories. For instance, Antoine Castex, Principal Data & AI Architect at L'Oreal, noted that the feature enabled them to label their extensive creative asset repository 100 times faster at a fraction of the cost of existing pipelines.

The service also integrates with BigQuery, allowing data teams to use the extracted context directly in SQL queries, simplifying multimodal analytics workflows. While still in preview, this product move indicates that cloud infrastructure vendors are attempting to accelerate the practical adoption of enterprise AI agents by lowering both the technical barriers and economic costs of data preparation.

Sources:Google Cloud Blog

Research

AWS launches USD billing for AI agents buying x402 products on Marketplace

Material
2026-11-20 07:02 GMT+8

Amazon Web Services has extended its Amazon Bedrock AgentCore payments capability to support direct purchases from AWS Marketplace using standard USD billing. Previously, when AgentCore payments became generally available in August 2026, agents could only transact over open protocols like x402 using stablecoin wallets. The new integration allows developers to configure agents to discover and buy x402-enabled software, APIs, data, and content, with charges settling directly on the agent developer’s AWS bill.

The mechanism operates through a specific flow: when an agent requests a resource from an AWS Marketplace seller, the seller responds with an HTTP 402 Payment Required status and a signed quote. The agent forwards this to AgentCore payments, which checks the session budget and issues a payment token if funds remain. The agent then retries the request with the token, and the transaction is recorded as a line item on the developer's AWS account. This process requires no pre-existing API keys or cryptocurrency wallets for the agent itself.

For enterprise deployment, this change addresses key governance concerns. Developers can set strict spending limits per session and restrict approved sellers using AgentCore policies. All transactions are auditable through AWS CloudTrail and visible in AWS Cost Explorer. Because both the buyer and seller operate within AWS accounts, the system meets traditional enterprise security and compliance expectations for procurement. At launch, the integration supports the x402 'exact' scheme, where sellers quote fixed prices per request.

Sources:Amazon Web Services

Research

AMD Shares Plunge After Short-Seller Report Flags China Exposure

Material
2026-11-20 06:31 GMT+8

Shares of U.S. chipmaker AMD plunged on Thursday, November 19, 2026, after short-seller Hunterbrook Research published a report flagging the company's heavy reliance on Chinese markets and supply chains. AMD rejected the report as a misrepresentation of its "business practices and relationships."

The release comes less than two weeks after former Nvidia CEO Jensen Huang admitted to an illicit relationship with a woman suspected of spying for Chinese intelligence, intensifying scrutiny of U.S. chip sales to China. The Hunterbrook report also cited AMD's financing arrangements with its own customers, which other analysts have warned could create an artificial impression that risks a large correction.

"If the smoke surrounding Nvidia becomes an inferno for the AI buildout, AMD will be engulfed by the flames," the report stated.

Sources:semafor.com

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

Regarding open-source integration, Dynamo has connected to existing reinforcement learning training stacks such as verl, NeMo RL, Slime, and prime-rl.