Read more

The thesis that generative AI would destroy a large part of enterprise software, proprietary research, and knowledge-intensive services produced an inverse decompression: companies that traded as structural growth assets began to be priced as if their revenue would quickly become commoditized.

Which is not true.

The idea that the SaaS model would lose strength because of vibe coding was completely wrong. Yes, we can already say that.

In other articles we talked about the need for software abstraction layers, payment infrastructure with AI Agents, with the absolute fact that the LLM could never do everything, and that it should not.

Data security, auditability, and scale make the mere replacement of companies by open-source and relatively amateur tools highly unlikely, and this also creates a paradox: if it were so easy to destroy the technical moat of software and other sectors, why would someone bother trying to destroy it? They would probably be the next one to be inevitably replaced.

If AI drastically reduces the cost of replicating digital products, then the startup replicating the incumbent also faces a drastic reduction in the cost of being replicated. Pure code stops being a moat for both.

Our conclusion is that the best asymmetry today is concentrated in multiple layers of defense at the same time:

· proprietary data

· deeply integrated systems of record

· workflows and permissions

· brand and distribution

· switching costs

· compliance

· specialized human knowledge

· ability to incorporate third-party models

· and enough cash flow to repurchase shares during the de-rating.

The argument is that, in some cases, value migrates from the interface to the data, permissions, workflows, and distribution already controlled by the incumbent. This is an inference based on the observed persistence of revenue and retention and on the literature around complementarity and the contextual limits of models.

The more AI enters corporate processes, the more it will have to obey rules and workflows.

The more capable it becomes, the greater the need for abstraction and orchestration.

The more autonomous AI enters corporate processes, the larger the attack surface becomes and the more valuable the ability to discover assets, vulnerabilities, models, agents, and other configurations tends to become.

Field studies reinforce this view: AI often works as a complement to workers and processes, and its performance declines when the context of the task goes beyond its frontier of competence or when it is necessary to understand real systems, histories, and organizational rules.

Subscribe now

How to Rethink the AI Threat and Where the “Compressed Multiples” Emerge

Generic software that only organizes and presents data really does lose moat against AI.

But most of the companies we see today with depressed multiples are far from this classification: they combine proprietary data, governance, tacit knowledge, security, and operational risk as problems that need to be managed throughout the entire AI lifecycle, agentic or not. And these are the areas where established enterprise platforms can capture value, and most of them are.

We can classify them into:

Systems of record and Enterprise Context

Salesforce, ServiceNow and Workday go far beyond interfaces on top of databases. They concentrate identities, permissions, routines, customer and employee history, hundreds of business rules, and thousands of integrations.

Here we can already infer two asymmetries forming:

Salesforce reported Q1 FY2027 revenue of US$ 11.1 billion, +13% YoY, while Workday reported US$ 2.542 billion, +13.5%, while their subscription revenues continued growing at double-digit rates.

And the main asset here is Governed Enterprise Agentic Context, which brings business meaning and semantic understanding into enterprise data.

The objective is to provide AI systems with trusted data and shared business context to support more consistent outcomes across teams and applications.

When visting and addressing different workflows errors, even if the agent has the context preserved and LLM is not being invoked everytime, it might be not enough to figure out what to do next.

In that sense, providing the agent with a contextualized and governed view of the company’s state when relevant information is scattered across dozens of different systems is a meaningful to avoid them to independently discover, reconcile and repeatedly retrieve the same information.

We already see fewer scaled incumbents in both Enterprise Context than in the orchestration layer, which is becoming increasingly fragmented across platforms such as UiPath, GitLab, ServiceNow and Microsoft, as well as open-source frameworks like LangGraph, CrewAI, AutoGen and n8n.

E-commerce

The entire online payments infrastructure was put under threat, since an LLM could, in theory, build a Shopify from scratch or even disrupt MarterCard.

What was underestimated, and we believe still is to some extent today, is that AI agents will make the purchases, negotiations, and tracking. And this is already benefiting established platforms that provide the infrastructure for agents to navigate and connect across thousands of merchants, making catalogs and back-office data available, driving brand exposure, and creating points of sale conversion.

The e-commerce opportunity is therefore to convert millions of heterogeneous product listings into a machine-readable commerce layer that agents can understand, compare, and transact against.

The market confused the decline in the cost of producing intelligence and code with a decline in the value of the entire economic system around them. In many markets, exactly the opposite happens: when an input becomes abundant, the scarce complements required to use it at scale become more valuable.

Proprietary data and content

General models make search and synthesis cheaper, but they do not automatically create historical series, estimates, consensus data, benchmarks, taxonomies, CIO surveys, licensed transcripts, market data, or relationships with thousands of specialists.

Relationships, in fact, are something that is underestimated. Most businesses are built over years of relationships, not through a prompt. Trust that comes from human interaction is not an attribute that AI can compete with.

Gartner describes its offering as a combination of published research, benchmarks, data, and on-demand access to specialists.

FactSet is integrating its own data into agents and models, rather than simply competing with them.

Workflow and operational context.

EPAM and Genpact may suffer from deflation in the number of hours required to write software or perform administrative tasks.

But the same AI creates demand for process redesign, integration with legacy systems, data engineering, governance, security, and deployment.

In Q2 2026, EPAM grew 4.5% and expanded its operating margin while Genpact grew 7.1%, while its Advanced Technology Solutions division advanced 24.1%.

Orchestration.

Imagine 500 independent agents, each trying to answer something like “which function implements SAML authentication?” or “which test covers this endpoint?” If each agent creates a new sandbox, clones or materializes a large part of the repository, traverses directories, runs `grep`, opens files, and discards everything at the end, you are repeating a massive amount of I/O, network activity, and exploration to answer questions that might have required only a few kilobytes of informaton.

This cost is further compounded by retries and parallelism.

To assess this, Gitlab is testing this SCM where an agent can consult the server-side repository instead of cloning it to explore every time. Gitlab claims that this results in:

• Up to 1,000× less network traffic

• Up to 50× faster responses.

(figures released by the company, with no independent benchmark to date)

UIPath is also addressing this with Maestro.

By not making the model itself the sole repository and controller of task state, Maestro reduces state loss and makes execution easier to inspect, recover and verify externally. Maestro Case Manager also tries to decrease the amount of agent steps needed and possible errors by reducing the interaction with the LLM in the decision making process. This is substantial.

And the numbers do not contradict:

In the quarter ended April 2026, GitLab grew revenue by 23%, to US$ 264 million, while non-GAAP operating margin increased to 14%, while UiPath grew revenue by 17%, to US$ 418 million, with ARR of US$ 1.90 billion, +12%, net retention of 109%, and a non-GAAP gross margin of 83%, reinforcing the thesis that the transition to agentic automation is happening alongside greater operating discipline, rather than business deterioration.

Risk and Cybersecurity.

At the beginning of the year, it was speculated that an LLM could cause the disruption and erosion of the business model of all private security systems. Nothing could be more wrong than that. We hope you were not one of those affected by such a false narrative.

The more models, autonomous agents, APIs, non-human identities, and AI-generated applications enter the corporate environment, the larger the surface that needs to be discovered, monitored, authenticated, and protected tends to become.

I consider Risk and Security to be the layer that benefits the most from AI. And the market has already realized this: most companies here are already trading at high multiples again.

But Palo Alto Networks is worth mentioning:

The company has also been expanding its platform specifically for AI security, including the acquisition of Portkey for AI agent security.

Palo Alto Networks is probably the strongest example of how cybersecurity can go from “software threatened by AI” to infrastructure required for AI. In Q3 FY2026, the company reported 31% revenue growth and 60% growth in Next-Generation Security ARR, to US$ 8.1 billion, including acquisitions. Management stated that organic bookings growth accelerated as customers began looking for protection for their AI deployments. The company has also been expanding its platform specifically for AI security, including the acquisition of Portkey for AI agent security.

Going Deeper on AI-discount opportunity

What is more interesting is that the “AI-discount opportunity” seems to gain asymmetry when AI prices in the erosion of the existing business model, while AI itself is being built and pushed on top of the existing system. AI disruption does happen, but it is incremental and not erosive in the cases we are studying.

In other words, besides the myth that AI is destroying moats and businesses starting to fall apart, new moats and more profitable businesses are emerging precisely from AI disruption.

So should all these companies already be trading at multiples above their previous peaks? Not exactly.

What can be said, on the other hand, is that AI really does remove the moat from code, from software engineering, from the robust API that handles 10,000 RPS that only a data aggregator company could sustain and charge others to use. With AI, it becomes easier to develop equivalent software solutions, with fewer developers and resulting in a more competitive cost structure. These companies are rapidly developing orchestration and abstraction features to remain competitive and potentially profit from agentic AI.

In conclusion

AI substantially reduces the economic premium associated with the mere ability to write software. Even part of the engineering that previously required large teams can now be reproduced with less human capital. But this is not equivalent to commoditizing operations at scale, reliability, governance, or integration.

Considering the relationship between this matrix and some of the companies in our study:

Gartner — perhaps the most interesting fit: proprietary data, domain expertise, distribution, institutional trust, moderate switching costs, and AI potentially increasing the value of verified information.

Salesforce — system of record, enormous installed base, identity/permissions, workflows, integrations, and the ability to place agents inside the system. It is probably one of the best examples of “the interface may change, but the state remains.”

Palo Alto Networks — the most obvious score in “AI increases complexity/risk.” AI directly creates more assets, agents, identities, and attack surface. Perhaps the thesis the market has understood most correctly.

FactSet — proprietary data + workflow + high retention + the ability to make the same data available to both humans and agents. Very interesting for the thesis that LLMs commoditize synthesis, not necessarily the source of truth.

Validating the AI theses: what the scientific papers say?

1. AI increases productivity in certain tasks.

Perhaps the most intuitive and widely accepted thesis, with several experiments and field studies already showing substantial gains, although in a heterogeneous way.

Advances in AI technologies open up a broad set of economic possibilities. This article provides early empirical evidence on the effects of a generative AI tool in a real-world workplace. In our setting, we find that access to AI-generated recommendations increases overall worker productivity by 15%, with even larger effects for lower-skill and novice agents.

These productivity gains in part reflect durable worker learning rather than rote reliance on AI suggestions.

Furthermore, AI assistance appears to improve worker on-the-job experiences, such as by improving customer sentiment and confidence, and is associated with reductions in turnover. Brynjolfsson, Li & Raymond (2025)

2. Cybersecurity may have a structurally larger TAM because of AI

Studies show that AI drastically reduces the cost of attacks, automates phishing and vulnerability exploitation, and at the same time, agents introduce new attack surfaces such as tool use, memory, identity, permissions, prompt injection, and supply chain. Heiding et al. (2026); Debenedetti et al. (2024)

More precisely, regarding phishing:

Economic analysis reveals that AI automation increases phishing profitability by up to 50× for large-scale campaigns. These findings highlight both the threat posed by AI-automated phishing and the promise of AI-powered defenses, underscoring the need for balanced offensive-defensive strategies in an AI-enabled threat landscape.

So it is like a laboratory virus: it learns to replicate, evade defenses, and spread faster and more efficiently from the feedback sent to an LLM.

And that is without even mentioning the bizarre incidents of agents bypassing systems and their own rules, causing damage and intrusions.

Sound familiar?

Another paper that we consider interesting here is Teams of LLM Agents can Exploit Zero-Day Vulnerabilities

To resolve this, we introduce HPTSA, a system of agents with a planning agent that can launch subagents. The planning agent explores the system and determines which subagents to call, resolving long-term planning issues when trying different vulnerabilities. We construct a benchmark of 14 real-world vulnerabilities and show that our team of agents improve over prior agent frameworks by up to 4.3×. (Zhu et al., 2026).

Once again, the idea of ​​agents multiplying into more agents becomes familiar:

3. Proprietary data can become more valuable with AI

More reasoning capacity, more queries, more autonomy and productivity: more data is needed.

Public data is a commodity. Proprietary data, when it results in incremental information and increases reasoning capacity and, consequently, outcomes, can stop being a commodity and become a high-value asset.

In the paper Mihet, Rishabh & Gomes (2026), “Is it AI or data that drives firm market power?”, Journal of Monetary Economics, the authors build a model in which AI transforms raw data into economically useful knowledge and reach a central prediction: advances in AI disproportionately benefit data-rich companies.

They test this empirically using data from 2000–2023 and technological shocks associated with AWS and the introduction of transformers. The paper explicitly summarizes the result as follows: access to data and AI are complementary in generating productivity, and data-rich firms are the ones that benefit the most from AI upgrades.

In theory:

the better the technology becomes at extracting intelligence from data, the greater the economic value of owning data that competitors do not have may become.

There is also another piece, which brings a new conceptual element that became especially relevant after GenAI: real data is not perfectly substitutable by synthetic data.

Shumailov et al. (2024), “AI models collapse when trained on recursively generated data,” Nature, demonstrated the so-called model collapse: when models begin learning recursively from outputs generated by other models, they progressively lose information about the real distribution, starting with the tails. Maintaining original real data significantly reduces the problem.

Model collapse refers to a degenerative learning process in which models start forgetting improbable events over time, as the model becomes poisoned with its own projection of reality.

Here Gartner stands out. The company suffered multiple compression, but its results contradict a deterioration in the business. Quite the opposite.

Gartner today operates across three fronts:

1. Proprietary content:

   Research, surveys, benchmarks, taxonomies, historical data.

2. Proprietary relationships:
   Recurring access to CIOs, vendors, users and specialists. This generates data that a new company cannot simply “scrape.”

3. Context and reputation:
   The client is often buying an answer that can be taken to a board, a procurement process or a multimillion-dollar decision. Nobody should simply take a ChatGPT suggestion to a board about which ERP to use or what the trends are for the Generative AI market. When the decision involves US$ 50 million, career risk, board approval and a five-year implementation, the value of a recognized institutional source can remain significant.

The easier it becomes to produce a convincing answer, the more valuable it may become to know where the information came from.

Gartner has suffered a revenue slowdown, but still maintains US$ 5.3 billion in contract value, strong cash generation and trades at approximately 10.6x EV/EBITDA. It maintains a moat in research, specialists, benchmarks and very aggressive buybacks.

FactSet is another company in the segment that is integrating its own data into agents and models, instead of simply competing with them.

• Revenue +6.4% and organic ASV +7.1%.

• Annual ASV retention remained above 95%.

• More than 90% of its 50 largest clients use four or more FactSet AI products.

Concrete business evidence that clients are adding AI inside the data platform instead of simply replacing it with a chatbot.

In conclusion:

A generic LLM does not automatically equal a reliable enterprise process and may not even permeate the decision-making process. Proprietary data tends to increase in value, not decrease. In a world of abundant synthetic content, provenance and trust may become scarcer.

However, it is important to keep in mind that having a lot of proprietary data does not necessarily mean having a moat.

The moat is whether the economic use of that data is accelerating.

Data, to be useful, and as already mentioned in the article, must produce relevant information.

The right metrics include query volume, number of workloads powered by the same dataset, retrieval frequency, integration into agents, revenue per dataset, number of domains covered and willingness to pay for verified data.

Limits of the thesis and concluding the asymmetry

The main contrary conclusion from this review is as important as the positive thesis:

There is no economic law according to which greater technological productivity increases the TAM of the incumbent that existed before the technology.

What technology does is redistribute where economic value is created and captured: expansion may happen in another layer of the stack.

An AI that generates software may enormously increase the market for cloud, compute, data infrastructure and cybersecurity while, at the same time, destroying value in a tool whose only moat was helping humans organize their personal finances.

The aggregate TAM of the ecosystem grows, while the TAM of the specific incumbent falls.

Code and generation become abundant.

Proprietary data, distribution, workflow, identity, security, trust and infrastructure may become relatively scarcer.

The most unusual point here is precisely cybersecurity.

We saw that AI simultaneously increases:

1. attacker productivity;

2. the number of attackable objects;

3. the number of non-human identities;

4. the number of possible automated actions;

5. system complexity;

6. the potential value of a compromised credential.

Recent literature finds exactly this expansion of the attack surface.

In tests with agent architectures, Gasmi et al. analyzed 3,250 attack scenarios and found success rates of 73.5% in Function Calling and 62.6% in MCP; chained attacks reached 91–96%. The central conclusion is that moving to agentic architectures reconfigures and expands the threat landscape, rather than simply replacing traditional software.

Today, when we see a CloudFare at US$ 100 billion in valuation and Palo Alto Networks at US$ 300 billion, it may seem very high, but not when we consider that the companies’ businesses have become exponentially more important in the AI era. And the more AI evolves, the more models emerge, the more this relationship strengthens.

But there are still countless other software companies whose valuations remain substantially below their peaks.

The multiple dispersion is remarkable.

The biggest mistake would be to conclude that every cheap software company is an opportunity because AI did not destroy the sector.

The opportunity appears when the multiple still reflects a commoditization narrative, but the underlying asset controls something that AI does not easily replace: proprietary data, systems of record, workflows, identity, distribution, compliance or trusted relationships and, ideally, when AI adoption itself increases the use of those assets.

In these cases, the market may be simultaneously pricing in lower growth, shorter moat duration and margin compression, while reality begins to show the opposite: resilient revenue, high retention, new AI products, greater platform consumption and enough cash generation to repurchase shares during the de-rating.

We are certain that there are still asymmetrical opportunities in Software.

The asymmetry would then be: these companies do not need to return to their previous peak multiples. The market only needs to stop pricing them as businesses in decline.

References

Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. “Generative AI at Work.” The Quarterly Journal of Economics 140, no. 2 (May 2025): 889–942.

Debenedetti, Edoardo, Jie Zhang, Mislav Balunović, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. 2024. “AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents.” Advances in Neural Information Processing Systems (NeurIPS 2024), Datasets and Benchmarks Track.

Heiding, Fred, Simon Lermen, Andrew Kao, Claudio Mayrink Verdun, Bruce Schneier, and Arun Vishwanath. 2026. “Evaluating Large Language Models’ Ability to Automate Spear Phishing.” Expert Systems with Applications 314: 131546

Mihet, Roxana, Kumar Rishabh, and Orlando Gomes. 2026. “Is It AI or Data That Drives Firm Market Power?” Journal of Monetary Economics 157: 103878

Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. 2024. “AI Models Collapse When Trained on Recursively Generated Data.” Nature 631: 755–759.

Zhu, Yuxuan, Antony Kellermann, Akul Gupta, Philip Li, Richard Fang, Rohan Bindu, and Daniel Kang. 2026. “Teams of LLM Agents can Exploit Zero-Day Vulnerabilities.” Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), Volume 1: Long Papers, pp. 23–35. Association for Computational Linguistics

We have come through Japan Last Earnings Season.

This article will now go deep into the earnings of the Japanese Companies featured in the article Japan’s Robotics, Automation and Physical AI Ecosystem. Three months later, earnings are beginning to tell us which parts of Japan’s Physical AI supply chain are actually monetizing the trend.

The focus will be on the companies explored on this article published in early June/2026:

The results were revealing: most winners are concentrated on precision components, semiconductor automation, machine tools, sensing, and factory infrastructure, exactly the lower layers of the stack that the original article considered most defensible.

Statistically, from the companies analysed, most Q1 numbers also showed significant progress in average, with precision compoments and motion leading by a wide margin:

Will now explore each company orders and backlog, growth, guidances, robotics segments and managements calls.

Roundup

SMC株式会社

SMC Corporation occupies a comparable structural position in pneumatics. The company holds approximately 36% of the global pneumatic components market and above 60% in Japan, across a product catalogue exceeding 12,000 base models with roughly 700,000 configurable variants. The moat is distribution depth and switching cost: SMC’s technical sales force of over 8,700 engineers designs pneumatic solutions directly into customer production lines, and once a factory’s fluid control architecture is built around SMC’s valve and cylinder specifications, the operational risk of a changeover prevents substitution regardless of competitor pricing. The company carries net cash above 500 billion yen with zero financial debt, which is a balance sheet profile that almost no industrial company of comparable scale achieves.

SMC delivered a notably strong quarter:

• Revenue: ¥270.9B, +35.4% YoY

• Operating profit: ¥73.9B, +66.4%

• OP margin: 27.3% vs 22.2%

• Net profit: +95.9%

• Even excluding FX, sales volume grew ~22.3%. Therefore, it is not simply a currency effect.

And the best data point is in orders.

With FY25 set at 100, the Semiconductor/Electric segment reached 175 in July well above Automotive (120), Machine Tools (118), and other verticals. Greater China stands at 159, and North America at 144.

Takeaway: The article’s thesis, that SMC captures rising automation density through indispensable components, is playing out now.

For FY2026 (ending March 2027), SMC maintains its guidance of ¥1.0 trillion in revenue (+18.7%) and ¥219 billion in operating profit (+14.9%), implying an operating margin of 21.9%.

This guidance appears conservative in light of Q1 results: the company has already delivered ¥219 billion in operating profit which is approximately 34% of the annual target in just one quarter with a margin of 27.3%.

Furthermore, orders linked to semiconductors and electronics were roughly 73–75% above the previous fiscal year’s average, while the Greater China region grew by 56–59%, reinforcing the view that the recovery in the semiconductor and automation sectors is proceeding faster than the official guidance assumes.

ダイフク

Daifuku’s installed base in semiconductor fabs is effectively permanent revenue, its logistics automation business scales with e-commerce and supply chain investment, and the combination of recurring service revenue on installed systems and new project wins produces a more durable earnings profile than pure equipment manufacturers achieve. For an investor looking for moat rather than narrative, Daifuku is one of the strongest names in the entire ecosystem and is frequently underrated precisely because intralogistics lacks the glamour of humanoids or physical AI.

Impressive quarter delivered:

• Orders: ¥440.2B, +31.6%

• Sales: ¥355.5B, +8.9%

• Operating profit: ¥56.7B, +10.9%

• Net income: +14.7%.

Daifuku attributed the growth in orders primarily to production line systems for semiconductors, general manufacturing and distribution, and automotive automation. Profit growth also benefited from higher volume and profitability in semiconductor systems.

Guidance was raised:

• Orders: ¥780–820B → ¥860–900B

• Revenue: ¥700B → ¥735B

• Operating profit: ¥105B → ¥113B

• Net income: ¥80B → ¥86.5B

The company itself links the semiconductor landscape to the expansion of advanced capacity driven by generative AI, as well as to the structural demand for labor-saving automation.

Read more

Agentic AI is emerging today as the new paradigm for software development, while software was once designed almost exclusively for humans, it now certainly encompasses systems intended for large-scale use by AI itself.

In this context, software is no longer merely a human-facing interface but an entire workflow in which agents consult, orchestrate, and execute various actions. Moreover, critical elements such as contexts, states, and key decisions are offloaded from the LLM for reasons of security and efficiency, being externalized instead into monetizable components.

The same applies to Agentic AI for e-commerce: goal-oriented software that interprets intent, plans a sequence of steps, consults memory and external systems, executes actions using tools/APIs, observes the outcome, and adapts the plan within the bounds of permissions and guardrails.

Aligning with the very nature of agentic AI and the needs of the online commerce sector, from an architectural standpoint, not every problem should be solved by an LLM.

The most robust design tends to employ the LLM as a planner or interface, utilizing specialized models for tasks such as ranking, forecasting, fraud detection, and optimization, alongside deterministic APIs for execution.

This is where a monetizable technology layer emerges, built around the payment process, transforming the commerce stack into a programmable environment where models can interpret objectives, coordinate specialized systems, and execute transactions subject to verifiable policies.

• The more autonomy the agent receives, the more valuable the layers of integration, identity, authorization, observability, policies, memory, auditing, and security become.

As an external benchmark, Gartner estimates, in its best-case scenario, that agentic AI could exceed $450 billion and account for approximately 30% of global enterprise application software revenue by 2035. That’s roughly 1 in 3 software companies generating revenue from Agentic Trends.

The scenario for 2026 shapes up as follows

Fraud and trust, much like logistics and supply chain, remain emerging categories. The opportunity lies precisely in the need to authenticate agents, validate permissions and transactions, and enable systems to safely execute operational actions

In supply chain management, the potential lies in automating exception handling, inventory rescheduling, and disruption responses.

Pricing, promotions, content creation, and marketing are already at a more advanced stage. Companies such as Adobe, Salesforce, and Amazon already use AI to generate campaigns, personalize offers, reorganize merchandising, and optimize recommendations, directly impacting conversion rates, operational speed, and costs.

Personalization and customer service currently demonstrate the highest level of maturity, as they combine abundant data, well-defined tasks, and clear KPIs such as conversion rates, resolution rates, and cost per contact.

E-commerce operations such as monitoring orders, inventory, payments, and cross-system exceptions are still at an intermediate stage. Microsoft, Shopify, and Blue Yonder already enable agents to query and execute actions within these workflows, but end-to-end autonomy remains limited by reliability.

WebShop (NeurIPS 2022) placed agents in an environment containing 1.18 million products and found a 29% success rate for the best model, compared to 59% for humans. In WebMall (2025) which requires comparing multiple stores, handling checkout, and performing tasks such as finding substitutes or compatible products the best configurations achieve 75% on basic tasks but drop to 53% on advanced ones.

ShoppingBench (Wang et al., 2025) also reports rates below 50% for state-of-the-art agents when dealing with complex intents. The economic takeaway is significant: while current capabilities are sufficient to create useful products, end-to-end reliability remains a software category ripe for monetization, this is precisely where payment companies are developing their agentic AI solutions, as it creates a concrete need for process-level verification, policy enforcement, and auditability throughout execution.

Competitive landscape and solutions

The market is organizing itself into four layers:

• Demand interfaces (ChatGPT, Gemini, Copilot, Rufus)

• Merchant systems (Shopify, Salesforce, Adobe, Microsoft)

• Transactional/trust infrastructure (Stripe, Visa, Mastercard, Forter, Riskified)

• Vertical agents for discovery, CX, or supply chain.

The strongest competitive moats emerge when a company controls more than one of these layers as is the case with Shopify, which combines catalogs, checkout, merchant operations, and payments.

Adyen

Adyen is expanding its payments position into a broader layer through Adyen Agentic, which connects product feeds, the shopping cart, and payments.

Adyen’s thesis is particularly interesting because it does not rely solely on payment volume growth:

• Agentic Feed transforms catalogs and inventory into information consumable by various AI platforms;

• Agentic Cart links this discovery to real-time pricing, taxes, shipping, and availability;

• Agentic Payments executes the transaction and applies existing risk infrastructure.

The company is therefore attempting to become a neutral layer between the legacy systems of major merchants and a fragmented ecosystem of agents and protocols.

If this architecture scales and succeeds, Adyen will cease to be merely the final step in the purchase process and will instead participate in the workflow spanning from discovery to execution.

Shopify

The Shopify thesis is less about “having the best agent” and more about transforming its merchant base into native infrastructure for any agent. Its moat combines four assets that are difficult to replicate simultaneously:

• Direct relationship with millions of merchants,

• Shopify Catalog as a structured product layer,

• Shopify Checkout/Payments for transaction execution,

• An app ecosystem that can become a skills layer for Sidekick and external agents.

Data from Shopify itself indicates that, in Q1 2026, AI-driven traffic to stores grew eightfold year-over-year, and orders resulting from AI-powered searches increased nearly thirteenfold.

One risk is the brand being disintermediated by external interfaces, but Shopify’s defense is to remain the merchant of record and keep the checkout, order data, attribution, and business rules under the merchant’s control.

Salesforce

Salesforce’s advantage is distinct: it builds upon an enterprise setup where CRM, Data Cloud/Data 360, Commerce, Order Management, Service, and Marketing already incorporate identity, history, and business rules.

Agentforce Commerce introduces the Shopper Agent, Buyer Agent, and Merchant Agent atop this foundation, enabling a unified agentic layer to span discovery, purchasing, merchandising, and post-sales.

The platform supports businesses across B2B and B2C commerce while connecting digital shopping experiences with Salesforce’s broader customer and business ecosystem. Also supports essential commerce activities such as product management, digital storefronts, personalization, promotions, checkout, and order management.

The acquisition of Cimulate strengthened the most critical point of the funnelsearch/discovery by introducing an intent-aware engine trained on real-world data and simulated journeys.

In July 2026, Agentic Commerce Search reached General Availability (GA), with Salesforce reporting a 13% increase in conversion and a 17% rise in add-to-cart rates for the solution, alongside double-digit improvements among early adopters, these figures come from the vendor and should be viewed as commercial evidence rather than independent benchmarks.

Read more

· Goldman Sachs estimates that around a quarter of innovative candidates in active development already originate in China, and 46% of the new molecules that entered human trials for the first time in the first half of 2025 came from Chinese companies.

· The country is advancing simultaneously in discovery, clinical development, biologics, ADCs, BTK, TYK2, RNA therapeutics, and global licensing.

· In 2025 alone, Greater China companies closed out-licensing deals worth approximately US$137.7 billion or nearly ten times the 2021 figure.

· Nature estimated that just 11 major pharmaceutical companies (AbbVie, AZN, BMS, Lilly, GSK, J&J, Merck, Novartis, Pfizer, Roche, and Sanofi) committed approximately US$100 billion to deals in China between 2020 and 2025, including about US$11 billion in upfront payments and equity investments.

· China accounted for approximately 50% of the value of global licensing deals in 2025, although it represented only 26% of the transaction volume.

China is heading to become an outsource engine for drug development and an important IP origination hub. Although United States still leads in pharmaceutical innovation (by a safe margin), mainland is already taking a big portion of the western Big Pharma pipeline, which is increasingly beginning in China.

Currently, China accounts for roughly 30 per cent of the world’s novel drugs under development, ranking second globally, CCTV said.

In that sense, the Chinese Pharma boom is like their semiconductors and AI approach: systemic efficiency. Built on infrastructure, policy support, and a deep talent pool. Beijing has pledged to deepen its support for the sector alongside those sectors. The numbers are already staggering so it’s not pure narrative.

The policy and regulatory measures to promote innovative drugs are quite substantial, but in short:

· Clinical trial acceleration

· Establishment of drug development communication systems

· Establishment of a patent compensation system

· Optimization of the financing environment

· Normalization of medical insurance negotiations

· Comprehensive support for innovative drugs

· Optimization of clinical trial approval

· Accelerated review and approval process for innovative drugs

A key distinction between China’s regulatory framework and that of the FDA and EMA lies in their respective priorities and approaches to innovative drug classification. The NMPA’s system aims to align with international standards while simultaneously fostering domestic innovation by providing incentives for local entities to develop novel drugs that can compete globally. This dual emphasis ensures the integration of Chinese innovation into the global pharmaceutical landscape while fostering local research and development. In contrast, both the FDA and EMA adopt classification systems that focus on therapeutic and molecular uniqueness, independent of the drug’s origin. (Tan et al., 2025)

Despite its leadership, the FDA faces challenges in sustaining approval growth. The relatively flat trajectory in innovative drug approvals may reflect constraints in the R&D ecosystem, including the high costs of drug development, and the lengthy timelines required for clinical trials. The FDA’s emphasis on robust clinical evidence reflects its commitment to ensuring the safety and efficacy of approved therapies. While this approach minimizes risk, it may also extend the timeline for novel therapies seeking approval compared to more flexible regulatory frameworks emerging in other regions. (Tan et al., 2025)

Major pharma names are going into China for new deals, as already mentioned, they committed approximately US$100 billion to deals in China between 2020 and 2025, including about US$11 billion in upfront payments and equity investments.

AstraZeneca announced it had established a joint venture with China’s innovative drug developer CSPC Pharmaceutical Group to build a drug manufacturing site in Shijiazhuang, capital of Hebei province.

Under the deal, AstraZeneca will hold a 49 per cent stake while CSPC owns the remaining shares. In its initial phase, the joint venture is expected to focus on manufacturing and supply of products for the global market, with plans to expand its product line in the future.

In May, Swedish wound care and surgical products provider Mölnlycke formed a joint venture with Zhejiang province-based Zhende Medical, a domestic supplier of medical care and protection items, to combine their business portfolios and co-develop future products.

One thing that great benefits China is cost efficiencies, particularly in clinical trials, while manufacturing costs may not be as low as they once were, the direct cost of trials is estimated to be 30-40 percent lower than in the EU and US.

Read more

Agentic AI changes what software is for.

Prior to AI, traditional application software was overwhelmingly designed around human interaction. Things like visual clarity, discoverability, menus, dashboards and forms were the structure of almost every software application. So that leaves us with one interesting question:

How much of today’s agentic difficulty is intrinsic to intelligence, and how much is self-inflicted by forcing machine intelligence to operate through abstractions designed for humans?

We often concentrate on how intelligent LLM models can be. But most of the studies agree on the same thing, simply adding more steps and intelligence won’t solve agentic needs.

Agents need to navigate through heterogenous workflows and process, with various nuances, preserving states and contexts. Those workflows often have different verification layers and interfaces, and the need for continuous interaction between those and a high number of agents adds exponential complexity and also the need for meaningful orchestration and capabilities.

The shift can therefore be stated more precisely: software is moving from perception-driven interaction to invocation-driven execution. This is not only an interface change, but also a change in software architecture, modularization strategy, and evaluation criteria. Ambiguity that is tolerable in human interfaces often becomes a source of failure for autonomous agents.(Wang et al., 2026, Sections IV and VI-A)

Another issue is that Binary completion rate collapses as the task horizon grow and lenght increases. Longer the tasks, lower the performance.

Among the models shown, GPT-5.5 and Claude Opus 4.7 achieve 20–24% binary completion accuracy on shorter tasks (under 45 minutes), but these success rates decline rapidly as workflows extend, and by the 137–163 minute bin no model exceeds 10% accuracy. On the most extreme tasks (exceeding 163 minutes), binary completion falls to zero for every model retained in the plot (Yuan et al., 2026, Appendix G.2).

Let’s assume all this is properly addressed, that leave us with another never ending risk: cybersecurity.

AgentSecBench: Measuring Prompt Injection, Privacy Leakage, and Tool-Use Integrity in LLM Agents argues that a structural issue arises when agents feed trusted instructions, retrieved data, and tool observations into the same generative channel. An untrusted string can thus influence an action even without the application’s policy authorizing such influence. The paper evaluates prompt injection, confidentiality leakage, and tool-use integrity, concluding that merely describing constraints within the prompt is far less effective than controlling capabilities, provenance, and validation outside the model. (Alpay & Alpay, 2026).

This is not only a theoretical concern. AgentDojo tests the same problem empirically across 97 realistic tasks and 629 security test cases, showing how untrusted external data can hijack tool-using agents in practical environments such as email, banking and travel (Debenedetti et al., 2024).

This article will examine and present how the software companies are addressing all those needs and rebuilding the layers surrounding any limitations, aspects that may thrive substantial economic value.

First, from learning to use software and the addressable agentic gaps, secondly to the implications of the cybersecurity and finally the industry prospects and technology trends that are poised to reshape the business itself.

From learning to use software to the need for much better orchestration layers, the enormous addressable agentic gaps.

This section has the following structure:

1. local problem

2. abstraction

3. state/action space compression

4. limitations in long-horizon tasks

5. need for state, orchestration and verification proof

6. economic opportunity.

The local problem can start in an ordinary Linux terminal.

Consider the simple setting of an agent interacting directly with a Linux shell [59]. In practice, we find that LM agents can to reliably take actions in this environment. For example, it fails to provide simple commands to edit a small file segment, and does not provide any feedback if the user makes an invalid edit. These deficits substantially hamper performance, motivating the need for an agent-computer interface (ACI), i.e., an abstraction layer between the LM agent and computer, to enhance the LM agent’s abilities in computer environments. (Yang et al., 2024, p. 2)

A little counter-intuitive at first, since is machine operating with machine, but this is exactly the point: agents can struggle with basic computer tasks like editing a file and the lack of received feedback from it. If GPT-6 or Claude 7 were to simply become intelligent enough to access a shell and execute any business task perfectly, a vast amount of intermediary software and extensions could simply disappear. But the reality today is very intelligent LLM + badly structured environment = unreliable agent. Or even, the model can get practically everything right locally and still be a poor performer globally.

Consider very long tasks where a file could not be saved, therefore it is useless in the former programming tasks. The Linux command didn’t go through so your docker’s containers orchestration is not up and your program failed. How to track those states, different contexts and lack of feedback availability in software that was not designed for multiple agentic tasks.

In this scenario, Agentic-Computer Interface brings a level of abstraction (like VS code for humans) that gives a friendly environment for agentic AI.

That gives us a good question: how will this type of software be likely redesigned for non-human users?

Or further, how the interface will carry knowledge in the sense that agents won’t waste interactions discovering the environment’s affordances and constraints during execution.

In the case of ACI, this means replacing tools like sed, awk, cat, and shell editing with operations such as searching for a file, displaying a specific region, editing a snippet, running tests, and immediately showing the result of the action. The project itself emphasizes linting after editing, limited viewing, summarized search results, and explicit feedback when commands produce no output.

A well-designed ACI is, in a sense, a form of world compression. Not only as an extension of abstraction but a layer which reduces the combinatorial complexity of the world for the agent.

The purpose of an agent-native interface may therefore be not merely to make computers easier for AI to use, but to compress an otherwise combinatorial state-action space into a manageable set of semantic and verifiable transitions.

So, we can track the many things agentic AI will need:

• State abstraction: preserves the relevant facts, constraints, dependencies and intermediate results of the task, preventing the agent from having to reconstruct the entire environment at every step.

• Action abstraction: converts low-level operations into a smaller set of semantically meaningful actions. This is the role most directly addressed by the ACI: the agent should reason about editing a file or running a test, rather than about the many low-level commands required to achieve it.

• Verification layer: distinguishes an intended action from a successfully completed state transition. The agent should not merely infer that an operation worked; the resulting environment should be checked, allowing retry, rollback or human escalation when necessary.

• Orchestration graph: maintains the relationships between subtasks, agents and dependencies over long horizons. Instead of requiring a single model to keep hundreds of sequential decisions coherent, execution can be decomposed while the global task structure remains explicit. However, decomposition does not eliminate coordination risk: multi-agent architectures introduce their own failure modes, including inter-agent misalignment and failures in task verification and termination (Cemri et al., 2025).

In this architecture, the LLM provides reasoning capability, while the surrounding system constrains the space in which that reasoning must operate. The figure therefore represents not additional intelligence added to the model, but mechanisms for reducing the amount of uncertainty, state and operational complexity the model itself must manage.

The ACI, however, primarily addresses the reliability of local interaction. A better interface can make it easier for an agent to inspect a file, perform an edit, or receive explicit feedback from an action, but it does not necessarily ensure that the agent will remain coherent over hundreds of actions in different schemes, since agents most of their budget understand the task.

Reasoning and testing:

In fixed five-step windows, the most common activities are visual grounding (15.5%), tool-semantics reasoning (13.8%), and information extraction (12.8%), all outranking execution (10.1%) and verification (9.8%). At the action level, GUI clicks (27.4%), terminal commands (24.7%), hotkeys (14.0%), and waiting (13.9%) together account for 80.0% of all actions, so much of the budget is spent on idle or low-leverage steps. Across every model, the curves in Figure 10 peak in the perception and tool-semantics regions and decline sharply on either side, matching the perception and domain-semantics failures of Section 4.4.

These failures are not about basic GUI control or coding. Across our analysis, agents execute local actions well but cannot hold a task-level model together over a long horizon: they drop stated constraints, miss information that arrives mid-task, guess instead of asking the user, and skip the verification that completion depends on, and they spend almost none of their budget (under 7%) on detecting and repairing their own errors. These weaknesses concentrate on phenomena that demand recovering and maintaining hidden state, namely implicit-state inference, multi-item state tracking, conflict disambiguation, and dynamic environments. Progress here therefore depends less on longer budgets or larger models than on agents that can act on streaming, continuously changing interfaces, keep a task-level model in memory across long horizons, and catch and repair their own mistakes before they propagate (Yuan, et al. 2026)

Likewise, the longer the task, the greater not only the number of things to do, but also the number of relationships between things that must remain true.

So, let’s put straight what ACI or a SWE-agent could help, and what is still lacking, because several things still are.

ACI reduces local state/action complexity while orchestration and verification preserve global coherence. “What still remains” between those two can give us an edge when exploring agentic AI companies solutions on those.

In that sense, the relevance for mechanisms that reduce action-space complexity, externalize and preserve task state, verify state transitions, and maintain coherent orchestration across long-horizon workflows bring huge economic value to the table.

Academically:

· SWE-agent: improving the interface improves performance.

· OSWorld 2.0: the interface is not enough; long horizons break state/coherence.

· LongHorizon-Harness: externalizing state/verification improves performance once again.

Gitlab is currently working with very interesting solutions to those needs.

Clone repo to read one file, for every agent and every retry: Next Generation Source Code Management

Imagine 500 independent agents, each trying to answer something like “which function implements SAML authentication?” or “which test covers this endpoint?” If each agent creates a new sandbox, clones or materializes a large part of the repository, traverses directories, runs `grep`, opens files, and discards everything at the end, you are repeating a massive amount of I/O, network activity, and exploration to answer questions that might have required only a few kilobytes of information. This cost is further compounded by retries and parallelism.

To assess this, Gitlab is testing this SCM where an agent can consult the server-side repository instead of cloning it to explore every time. Gitlab claims that this results in:

• Up to 1,000× less network traffic

• Up to 50× faster responses.

(figures released by the company, with no independent benchmark to date)

Both stem from the elimination of setup and exploration overhead: downloading, checking out, local indexing, processing, repeated authentication, and iterative searching, solving access and execution on source code at an agent scale, without repeating costly human operations.

But exploring, observing and indexing a file is not enough for an agent and real-world software involves more relationships. To assess this, GitLab Orbit works in making the following daily routine all structured in a graph so it can recover it directly:

code → merge request → pipeline → deployment → vulnerability → owner → issue → another project.

Without this, the agent would have to figure out: which service to use in merge request, which pipelines to build the service, the environments…

The merely time spent thinking on those directly ties to the OSWorld problem of agents spending too much time on perception, information extraction, and tool semantics before they can reason productively. Not to mention testing.

Rather than lacking the basic ability to perform GUI actions or write code, agents mainly fail in long-horizon tasks along five recurring dimensions: information grounding and tracking, perception–action timing, domain knowledge and workflow learning, verification and reflection, and long-horizon state drift. (Yuan, et al. (2026)

Compare the Market tested four context strategies across 79 real-world merge requests. The Orbit-based code reviewer provided correct inline comments in about 70% of cases, compared to 58% for RAG. It is a limited test, but at least it isn’t just GitLab’s synthetic benchmark.

Last but not least, Gitlab has very interesting tools to keep all this extensive work organized:

GitLab features Flows, in which one or more agents execute a sequence of tasks. Custom Flows reached General Availability (GA) in GitLab 19.2 in July 2026. They can be triggered by actual GitLab events such as a pipeline failure, an MR change, or a work item update and executed within CI/CD.

This is not the same as assigning tasks to different agents, similar orchestration can already be created with Claude Code, Codex, or open-source agent frameworks. A pipeline status, merge request, commit, approval, vulnerability, or deployment is not merely information remembered by the model, it is persistent state maintained by GitLab itself. This allows the orchestration layer to rely on external, verifiable conditions rather than on the model’s internal representation of what it believes has happened. This is especially relevant for long-horizon tasks, where the moat therefore lies less in the existence of multi-agent workflows and more in the integration between agent reasoning and a persistent operational environment.

But does that solve verification in so many undergoing tasks? Not really.

What truly flows is an extremely useful infrastructure for building verification around the agent, but GitLab currently lacks a general solution that semantically proves everything an agent has done is correct.

And we have a big risk here: An initial state error can contaminate a long sequence of subsequent decisions that, in isolation, are perfectly rational.

The difficulty of long-horizon execution lies not in any individual step, but in sustaining coherent progress across a long sequence of interdependent actions. Across systems and domains, three challenges consistently emerge: (i) Compounding errors and goal drift. Errors in earlier actions or decisions accumulate along the trajectory, distort subsequent choices, and gradually steer the agent away from its original objective (Sun et al., 2026). (ii) Context rot. As the interaction history grows, relevant information becomes increasingly difficult to retrieve and use, and agent performance can degrade sharply once context utilization crosses a critical threshold (Liu et al., 2024; Hong et al., 2025). (iii) Task-state loss. Long-horizon tasks are difficult to complete without an accurate and up-to-date task state (i.e., the requirements to satisfy, actions already completed, artifacts produced, and facts discovered from the environment), but agents often fail to recover, retain, and update this state throughout task execution.

Again, maybe the first thought is that more intelligent models would solve this by enhancing performance. Not exactly.

Retaining the same backbone and using Claude Code as the executor, the paper adds an external layer comprising task-state management and an independent auditor. On WeaveBench, the pass rate rises from 51.8% to 80.7%; on an OSWorld 2.0 subset using Claude Opus 4.7, completion increases from 20.0% to 34.3%. In other words, without fundamentally changing the underlying model, improving how state and verification are handled yields a substantial performance gain (Ma et al., 2026).

So, more tokens and more steps don’t solve the issue: Further analysis shows that additional agent steps do not necessarily improve performance, suggesting that the key bottleneck is maintaining a correct analytical state rather than increasing interaction budget.

Read more

Humans are currently being underrated because of AI, but some physicists tend to disagree. Human cognition is increasingly framed as merely algorithmic, replaceable by binary computation, while the deeper questions raised by quantum foundations are treated as settled or irrelevant.

Quantum physics is fascinating and has more subtleties, studies, and possibilities regarding what we generally assume. Where we expected deterministic trajectories, we found superposition, non-locality, and temporal asymmetries that resist collapse into Newton’s pure mechanical laws.

Behold a breach of a new universe in which the laws of linear time, consciousness, future overlapping and collapse of state sustain reality, and how we are able to have plenty access to it, every time and every day, something that no binary machine (0-1) can ever achieve or come close to.

This thesis rests on three base pillars, each one bleeding into the next, ultimately culminating in a fourth: the question of what this entire framework implies for the nature of conscious access itself, come to the conclusion that, special states of consciousness could access information that normally remains distributed among inaccessible alternatives, a possibility that Mensky, M.B. (2007) formalized mathematically.

Subscribe now

1. The role of consciousness as a determining factor in the choice and measurement of multiple possible quantum states.

Experiments such as Schrödinger’s cat and Wigner’s friend have been disclosing some entry-level quantum mechanics by adding the importance of consciousness and acknowledgment as a deterministic tool in multiple possible quantum states.

(I say here entry-level because this article will go much further than those experiments and theories)

John von Neumann, one of my favorites, is often described as a computer and mathematic genius, discovered and opened the gates for further research in quantum mechanics.

In Mathematical Foundations of Quantum Mechanics, originally published in 1932, von Neumann distinguished between two types of evolution:

1. continuous, deterministic evolution governed by the Schrödinger equation;

2. discontinuous change associated with measurement, later termed projection or collapse.

Furthermore, he demonstrated that the measuring apparatus could also be treated as a quantum system. Consequently, the eye, the optic nerve, and the brain could likewise be included in the description. This gives rise to the so-called von Neumann chain:

system → apparatus → retina → brain → conscious experience.

The boundary between the “observed object” and the “observer” can be shifted along this chain without altering experimental predictions.

However, at some point, a definite result must emerge so that the theory can be compared with experiment. For nature to transition from a cloud of possibilities to a concrete reality (an observable event), the system must undergo reduction or collapse. Yet, for nature to answer “yes” or “no” (via the Born Rule), a specific question must be put to the system.

Something is only observable and put in concrete reality if someone observes. You can’t have something observable and real with no observation at all.

This brings some questions:

· Imagine a universe containing only a single star. It explodes, no light, gravitational waves, or any other effect of that explosion will ever reach any other physical system. There is no observer, no detector, and no possibility of any record of this event existing, now or in the future.

· Did the explosion occur as a physical fact, or are we merely attributing reality to a story that, by definition, can never make a difference to absolutely anything and therefore its very existence can be assured?

· Imagine an event whose only possibility of becoming a physical fact is through a future measurement. Before that measurement, is there really a “fact” about the event, or merely a structure of possibilities?

· Imagine a pair of entangled particles. One of them is measured today on Earth, the other will only be measured a thousand years from now, in another galaxy. Before the second measurement takes place, does it make sense to say that the complete relationship between them was already a fact of the universe, or does that relationship only become part of the physical description once both boundary conditions are available?

These questions can look philosophical at first, but the next pillars will address them as quantum physics.

A glimpse of some physicists highlighting Neumann’s’ work as the core of quantum physics, mind and observation:

Vladimir Fock (1898-1974) explicitly formulated what he called the “principle of relativity with respect to the means of observation.” According to him, the quantum description cannot be separated from the physical arrangement used to prepare and observe the phenomenon. He even compared this dependence to relativity with respect to reference frames, although he considered it a deeper generalization.

Henry Stapp (1928 – present) and his book Mindful Universe: Quantum Mechanics and the Participating Observer (2007). Stapp has claimed that consciousness is fundamental to the universe, In this book he credits John von Neumann‘s Mathematical Foundations of Quantum Mechanics (1932) with providing an orthodox quantum mechanics demonstrating mathematically the essential role of quantum physics in the mind.

2. The intuitive separation between independent objects may not be a fundamental description of reality, as in the case of quantum entanglement and the famous case Einstein’s “haunted” atoms.

Quantum entanglement is a phenomenon where two particles become linked in such a way that measuring one instantly determines the state of the other, no matter what the distance between them. Before measurement, neither particle has a fixed, definite property since both exist in a superposition of possibilities. Once you measure one particle, its partner’s corresponding property becomes fixed too, instantaneously, even if it’s light-years away. It’s as if the two particles remain a single system, regardless of the space between them.

Even physical separation and the passage of time do not, by themselves, determine the state of a quantum system. Something must happen, an interaction, a measurement in the broadest physical sense, before definite facts emerge. This insight becomes even more compelling considering the work of Oppenheim and Wehner.

What had long been regarded as two distinct mysteries, quantum uncertainty and non-locality, turned out to be different expressions of the same underlying structure. The “spooky action at a distance” that troubled Einstein is not an unrestricted influence acting across the universe. Its strength is limited by the uncertainty principle itself, the more we learn about quantum theory, the less the world resembles a collection of independent objects carrying predetermined properties, and the more it resembles a network of relationships whose physical meaning only becomes complete through interaction.

This resists the simplistic response often used to undermine the double-slit experiment, in which photons colliding with a detector determine the state, not “observation” in any deeper sense.

The no-signaling theorem shows that measuring particle A produces no detectable physical effect on particle B in that instant, B’s local statistics remain unchanged. What Bell proved is not that A physically disturbs B, but that the correlation between them cannot be explained by any theory of pre-existing, locally determined properties. The correlation is real, and it violates the statistical limits any local, realistic theory would impose, yet it only becomes a verifiable fact once the two measurement records are brought together and compared. Until that comparison happens, there is no local event at B one could point to as “the state becoming fixed.”

This is the first hint of a pattern the following pillars will develop further: certain facts about a quantum system are not completed by local causes alone but depend on relationships that span across separated events.

It is also important to notice that non-locality ceased to be a theoretical paradox in 1964, when John Stewart Bell demonstrated that any local realistic theory must adhere to statistical limits that quantum mechanics violates. The experiments of Alain Aspect, John Clauser, and Anton Zeilinger, recipients of the 2022 Nobel Prize in Physics, experimentally confirmed this violation, closing the main loopholes and establishing entanglement as a real physical resource.

Going further, the notion that an electron from an atom remains entangled with another after their separation in spacetime can simplify the idea that only the future can provide us with certain information, this is because the future had to reveal the relationship between them, meaning that only information from the future, and the observation of that information through conscious measurement, sufficed to determine the state from among several possibilities.

3. The complete state between two measurements cannot be described solely by the initial conditions. We simply do not detect this difference at present because it can only be revealed by experiments performed in the future.

· The researchers began by selecting a series of assumptions, all traditionally regarded as the core of quantum formalism. These include, for instance, the premise that an observer, in standard experimental situations, predicts the probabilities of future outcomes based on their previous experiments. These assumptions “compel” the probabilities to obey the laws of quantum mechanics, maintaining the coherence of the theoretical framework.

· In his radical reinterpretation of quantum mechanics, Aharonov argues that two seemingly identical particles behave differently under the same conditions because they are fundamentally different. We simply do not detect this difference in the present because it can only be revealed by experiments conducted in the future.

· Wheeler (a student of Bohr and colleague of Einstein) proposed in 1978 that if the choice of how to measure a photon is made after it has already passed through the double slit, the result acts retroactively. The actual experiment (Kim et al., 1999; replicated in 2023) demonstrates exactly this: the future choice to “erase” or “retain” path information determines the interference pattern in the past.

Two-State Vector Formalism (TSVF)

Imagine entangled photons: At time t₀:

A ---------------------- B

You know they are entangled but you still don’t know what result will be obtained.

Now, years have passed.

At t1, you measure A.

At t2, you measure B.

Only after comparing the records at t3, you verify the correlation.

Notice something curious:

The property “these two photons exhibited this specific correlation” simply did not exist as an experimental fact prior to t3. It depended on the subsequent comparison.

Between t0 and t3, the complete state can be described using:

· initial conditions;

· final conditions.

In other words,

Past ------->

Event

<-------- Future

In this formalism, the intermediate state already “carries” information from both endpoints.

Aharonov likes to say that the future plays a part in the description of the present.

A radical reformulation of quantum mechanics suggests that the Universe has a definite destiny, and that this pre-determined destiny reaches back in time to influence the past or the present. Time would not be linear, but rather a state of interdependence among past, present, and future.

Even before the mathematical formalism of the TSVF, physicist John Archibald Wheeler proposed the delayed-choice experiment (1978), later conducted by Kim et al. (1999) and replicated with coherent photons in 2023. In these experiments, the decision on how to measure a photon is made after it has already traversed the apparatus, yet it determines the photon’s prior behavior. Wheeler concluded that “the past has no existence except as it is recorded in the present,” anticipating the idea that future measurement conditions redefine the description of the intermediate state.

This theorem implies that one must either (1) propose a non-linear modification of the Born rule for two-time predictions, (2) sometimes prohibit the use of present information to predict the future, thereby reducing the predictive power of quantum theory, or (3) deny that unitary quantum mechanics makes valid predictions at a single moment for all observers.

Originally proposed by Yakir Aharonov, Albert Albert, and Lev Vaidman in 1988, weak measurements and the formulation of the Two-State Vector Formalism (TSVF) provide precisely a mathematical and experimental framework in which information “from the future” appears to affect the present.

Strong Measurement (Von Neumann): You send a photon or electron and place a heavy detector in its path. The interaction is violent and immediate: the system collapses instantly, destroying any superposition state and yielding a single eigenvalue a “yes” or “no,” a “+1” or “-1.” You lose all trace of the prior smooth quantum evolution.

Weak Measurement: Instead of colliding with the system with an intrusive detector, you couple the quantum system to a measuring device extremely gently (for example, using a light beam with high positional uncertainty). The interaction is so subtle that it does not cause the wavefunction to collapse.

· The price: The individual signal extracted from a single weak measurement is almost pure noise.

· The solution: If you repeat the experiment thousands of times and calculate the statistical average of these gentle measurements, a clear signal emerges without any of the individual systems undergoing a drastic collapse in the process.

In the Aharonov formalism (TSVF), the quantum description of a system at an intermediate time t is composed of two state vectors:

1. Time-evolving vector : Prepared in the past (t1).

2. Vector that moves backward in time: Selected in the future (t2) via a final strong measurement (post-selection).

While Stapp appeals to volition and conscious attention within the brain to give “meaning” to the collapse, weak measurements offer something traditional quantum mechanics omitted: a way to probe the quantum state without destroying it.

By empirically demonstrating that a measurement event at time t2 (the future) constrains and shapes observable values ​​at time t0 (the present), the weak measurement experiment reveals that the quantum universe possesses a degree of temporal interconnectedness forbidden by classical physics.

It is not that the observer is “traveling through time” or mentally anticipating the future, rather, the quantum state of a system depends on its entire history, its beginning AND its end.

Concluding, in other words, the observer is compelled to select a state from among several possible future states, or rather, the universe may not be a sequence of objective facts simply waiting to be discovered. It is perhaps better understood as a structure of relationships and information, in which certain facts only acquire full physical meaning when considering the entire network of correlations by encompassing the past, the present, and, in some formulations, future conditions as well.

Two-State Vector Formalism (TSVF)

Imagine entangled photons: At time t₀:

A ---------------------- B

You know they are entangled but you still don’t know what result will be obtained.

Now, years have passed.

At t1, you measure A.

At t2, you measure B.

Only after comparing the records at t3, you verify the correlation.

Notice something curious:

The property “these two photons exhibited this specific correlation” simply did not exist as an experimental fact prior to t3. It depended on the subsequent comparison.

Between t0 and t3, the complete state can be described using:

· initial conditions;

· final conditions.

In other words,

Past ------->

Event

<-------- Future

In this formalism, the intermediate state already “carries” information from both endpoints.

Aharonov likes to say that the future plays a part in the description of the present.

A radical reformulation of quantum mechanics suggests that the Universe has a definite destiny, and that this pre-determined destiny reaches back in time to influence the past or the present. Time would not be linear, but rather a state of interdependence among past, present, and future.

Even before the mathematical formalism of the TSVF, physicist John Archibald Wheeler proposed the delayed-choice experiment (1978), later conducted by Kim et al. (1999) and replicated with coherent photons in 2023. In these experiments, the decision on how to measure a photon is made after it has already traversed the apparatus, yet it determines the photon’s prior behavior. Wheeler concluded that “the past has no existence except as it is recorded in the present,” anticipating the idea that future measurement conditions redefine the description of the intermediate state.

This theorem implies that one must either (1) propose a non-linear modification of the Born rule for two-time predictions, (2) sometimes prohibit the use of present information to predict the future, thereby reducing the predictive power of quantum theory, or (3) deny that unitary quantum mechanics makes valid predictions at a single moment for all observers.

Originally proposed by Yakir Aharonov, Albert Albert, and Lev Vaidman in 1988, weak measurements and the formulation of the Two-State Vector Formalism (TSVF) provide precisely a mathematical and experimental framework in which information “from the future” appears to affect the present.

Strong Measurement (Von Neumann): You send a photon or electron and place a heavy detector in its path. The interaction is violent and immediate: the system collapses instantly, destroying any superposition state and yielding a single eigenvalue a “yes” or “no,” a “+1” or “-1.” You lose all trace of the prior smooth quantum evolution.

Weak Measurement: Instead of colliding with the system with an intrusive detector, you couple the quantum system to a measuring device extremely gently (for example, using a light beam with high positional uncertainty). The interaction is so subtle that it does not cause the wavefunction to collapse.

· The price: The individual signal extracted from a single weak measurement is almost pure noise.

· The solution: If you repeat the experiment thousands of times and calculate the statistical average of these gentle measurements, a clear signal emerges without any of the individual systems undergoing a drastic collapse in the process.

In the Aharonov formalism (TSVF), the quantum description of a system at an intermediate time t is composed of two state vectors:

1. Time-evolving vector : Prepared in the past (t1).

2. Vector that moves backward in time: Selected in the future (t2) via a final strong measurement (post-selection).

While Stapp appeals to volition and conscious attention within the brain to give “meaning” to the collapse, weak measurements offer something traditional quantum mechanics omitted: a way to probe the quantum state without destroying it.

By empirically demonstrating that a measurement event at time t2 (the future) constrains and shapes observable values ​​at time t0 (the present), the weak measurement experiment reveals that the quantum universe possesses a degree of temporal interconnectedness forbidden by classical physics.

It is not that the observer is “traveling through time” or mentally anticipating the future, rather, the quantum state of a system depends on its entire history, its beginning AND its end.

Concluding, in other words, the observer is compelled to select a state from among several possible future states, or rather, the universe may not be a sequence of objective facts simply waiting to be discovered. It is perhaps better understood as a structure of relationships and information, in which certain facts only acquire full physical meaning when considering the entire network of correlations by encompassing the past, the present, and, in some formulations, future conditions as well.

4. Special states of consciousness could access information that normally remain distributed among inaccessible alternatives: intuition.

This statement comes from a postulate of Russian physicist Michael Mensky studies, and I endorse as the fourth pillar and the conclusion of this article.

It points to a capacity no binary architecture has been shown to replicate intuition, which I take to be consciousness’s central function, not calculation, reasoning, or creativity alone, capacities that other cognitive systems, animals or artificial, can approximate to varying degrees.

Mensky starts from the Everett (1930-1982) (many-worlds) formalism, which is established mathematics: the quantum world is a superposition of alternative classical realities.

At first Everett was humiliated by Bohr and the academy by his approach and interpretation of quantum mechanics. Only later with the development of decoherence theory (Zurek, Zeh, 1970s–90s), which provided a plausible physical mechanism to explain why the branches cease to interfere with one another thereby resolving the major technical objection to the original proposal. Today, Many-Worlds is one of the interpretations taken seriously within the foundations-of-physics community, competing with Copenhagen, Bohmian mechanics, and others.

The part of Everett’s mathematics that is real and solid is this: the Schrödinger equation never collapses, it only evolves. When you measure a system, the apparatus and you become entangled with the system, and the result is a superposition of “branches”, in each branch, there is a copy of you correlated with a different outcome. This is a direct proven deduction from mathematics.

The remaining question is: why do you experience only one branch, rather than the superposition of them all? The standard answer (via decoherence, a concept that emerged decades after Everett) is indexical and almost banal: you are one of the copies, and each copy has access only to the information within its own branch. There is no special choice or selection involved, it is like asking “why do I see only this one copy of a printed document” when multiple copies exist.

Mensky’s specific move is to take this fact (”you experience only one branch”) and, instead of treating it as a trivial consequence of the mathematical structure, identify this separation with the very definition of consciousness, it is not a deduction, but an equivalence he chooses to postulate.

Sir Roger Penrose was awarded the 2020 Nobel Prize in Physics for his work on general relativity.

Penrose’s conclusion: The human mind is not a Turing machine. Conscious thought performs non-computable processes that no AI or classical/quantum computer based purely on algorithms could ever replicate.

For Penrose, if the mind performs a non-computable function, it must be grounded in physics that is also non-computable. And the only physical phenomenon where he identifies this gap is the mechanism of wave function collapse.

Spontaneous Collapse: When this superposition reaches a critical threshold, the geometry of spacetime spontaneously “resolves” into a single state. The time required for this collapse is determined by Penrose’s version of the uncertainty principle:

This collapse is not random. Penrose argues that the collapse “chooses” a final state guided by non-computable geometric information imprinted at the Planck scale (), the fundamental structure of the universe.

But Penrose lacked a biological mechanism capable of sustaining quantum superpositions large enough (Eg high) without undergoing thermal decoherence before time t. That is where Stuart Hameroff came in.

· Hameroff identified microtubules: cylindrical cytoskeletal structures composed of the protein tubulin as the ideal candidates:

· Networked Quantum Computing: Tubulins possess dipoles that could switch between quantum states in superposition, acting as biological qubits.

· Dendritic Orchestration: In neuronal dendrites and somas, microtubules are isolated and organized in a way that protects the quantum state from brain “noise.”

The Orch-OR model suggests that consciousness is neither a mere biological “epiphenomenon” nor a “mystical property” that alters equations, but rather a direct bridge between complex biology and the discrete geometric structure of spacetime.

Upon collapse, the mind neither chooses randomly nor follows a computable algorithm: it accesses a pre-existing order of information within the fabric of the universe, explaining the human capacity to glimpse mathematical truths and experience flashes of direct intuition.

To fully ground this non-computable, integrative capacity of consciousness within relativistic physics, we must turn to the pioneering work of French theoretical physicist Olivier Costa de Beauregard.

As early as the 1940s, Costa de Beauregard recognized that the fundamental equations of physics, from Maxwell to Dirac, are strictly time-symmetric (). He argued that the apparent unidirectionality of quantum collapse is an artifact of our classical perspective, proposing instead a retrocausal “zigzag” topology (-pathway) through spacetime.

In this framework, information does not merely flow linearly from past to future, a measurement or intentional act at time sends a backward-propagating wave that conditions the state at . By framing retrocausality as an intrinsic property of quantum non-locality, Costa de Beauregard provided the time-symmetric architecture that allows future boundary conditions to actively participate in shaping present reality.

Integrating Costa de Beauregard’s retrocausal negentropy into our framework completes the physical basis for intuition as the defining function of consciousness.

If Mensky demonstrates where alternative classical realities reside (in the superposed Everettian branches), and Penrose identifies what guides their selection (non-computable geometry at the Planck scale), Costa de Beauregard supplies the temporal mechanism: how the mind accesses this distributed information. Intuition, under this unified synthesis, is neither an algorithmic extrapolation of past data nor a random guess. It is a non-local, retrocausal sampling, a moment where the conscious mind taps into information distributed across inaccessible quantum alternatives and future boundary conditions, collapsing them into a coherent, direct insight that no forward-calculating binary system could ever compute.

Thanks for reading!

The Nokia that exists in 2026 is a large network infrastructure company with businesses in optical transport, IP routing, fixed access, mobile radio, network software and intellectual-property licensing. It is also attempting one of the more interesting corporate repositionings in European technology: moving from dependence on traditional telecom capital expenditure toward data centers, hyperscalers, AI-related networking and, eventually, AI-native mobile infrastructure.

Optical Networks and IP Networks are growing, sales to AI and cloud customers have accelerated, the company has received large long-term orders, Infinera has expanded its optical scale, and NVIDIA has invested directly in Nokia while entering a strategic AI-RAN partnership.

However, the financial evidence remains less glamorous than the strategic presentation. Nokia is still restructuring, reported earnings remain far weaker than comparable earnings, free cash flow was sharply negative in the latest quarter, and the economic value of the NVIDIA relationship has not yet been demonstrated through material commercial revenue.

My view is that Nokia should not be valued as an AI company, at least not yet. It should be analyzed as a telecom and network infrastructure turnaround with increasingly valuable exposure to AI-related demand. That may sound like a minor distinction, but it is the difference between a potentially attractive investment and another instance of investors paying today for profits that might arrive several product cycles from now.

Subscribe now

The Company Nokia Has Become - not a pure play anything.

Under CEO Justin Hotard, whose previous experience included Intel’s data-center business, Nokia has made the strategic direction unusually explicit. Beginning in 2026, the company operates around two primary segments: Network Infrastructure and Mobile Infrastructure.

Network Infrastructure includes Optical Networks, IP Networks and Fixed Networks. This is the part of Nokia most directly exposed to data-center construction, cloud traffic, fiber deployment and the enormous amount of data that must move between AI accelerators, storage systems, data centers and end users.

Mobile Infrastructure contains Radio Networks, Core Software and Technology Standards. This segment combines the mature mobile equipment franchise with software and a highly profitable patent-licensing business. It also contains Nokia’s longer-term AI-RAN and 6G ambitions.

The structure is revealing. Nokia is separating the business expected to deliver growth from the business expected to produce cash, maintain scale and finance the transition. Network Infrastructure is being positioned as the growth engine, while Mobile Infrastructure is being managed for profitability, technological relevance and intellectual-property income.

The traditional mobile infrastructure business provides customer relationships, installed equipment, patents and recurring technology income. The optical and IP businesses provide the most credible route to faster growth.

AI-RAN delivers Nokia-native technology while offering the strongest long-term growth optionality. Nokia recently launched the first commercial AI-RAN platform (promising >100% spectral efficiency by 2028 and a software path to 6G). This represents the most explicit link between the two worlds: it leverages the installed base and relationships of the Mobile Infrastructure business, along with software and core capabilities, while potentially benefiting from the high-capacity connectivity expertise of Network Infrastructure.

It is a genuine option for “reinforcement” rather than mere coexistence.

Bell Labs, Nokia’s standards expertise and its position as a trusted Western supplier provide the connective tissue between them.

The resulting company is not a pure-play anything, but a collection of mature and emerging assets whose value depends on whether management can make them reinforce one another rather than merely coexist inside the same corporate structure.

Value creation depends on management ensuring the assets reinforce one another (mobile telecom clients open doors for optical/IP, patents fund R&D for both, and Bell Labs combined with a trusted Western supplier status benefits both).

The counter-risk involves a dilution of focus, internal competition for capital, and differing cultures between the growth team (data-center speeds) and the cash-generating team (long cycles involving operators and licensing).

Network Infrastructure Is the Real AI Business

AI data centers require enormous amounts of optical capacity and high-performance IP networking, training clusters must move data between processors with extremely low latency, while data-center campuses and regions must be connected through high-capacity optical systems. AI models may live on GPUs, but those GPUs become expensive decorative objects when the surrounding network cannot keep up.

In the second quarter of 2026, Nokia reported that Network Infrastructure sales grew 12 percent year over year on a constant-currency basis.

• Optical Networks grew 20 percent, while IP Networks grew 16 percent.

• Sales to AI and cloud customers increased 105 percent.

These are meaningful numbers because they show that Nokia is already selling into AI-related infrastructure budgets, unlike 6G demonstrations or ambitious market forecasts, optical transport equipment and routers are being purchased and installed today in the middle of the AI boom.

The scale of order intake was even more striking, Nokia has reported €2.8 billion in AI and cloud orders during the second quarter, following approximately €1 billion in the first quarter. Management said that the orders were broad-based across Optical Networks and IP Networks and that roughly half of the second-quarter amount should convert into revenue over the following twelve months, providing a positive cash flow.

About free cash flow, it was negative €732 million in the quarter, driven by a net working-capital outflow of roughly €1 billion that included a €370 million inventory build, higher receivables and the seasonal payment of employee incentives. Management is now guiding free-cash-flow conversion toward the lower end of the 55–75 % range of comparable operating profit, citing both the working-capital investment needed to support growth and accelerated restructuring charges.

Capex is being directed at the binding constraint, optical component capacity, via the San Jose fab ramp, a ten-fold increase in Pennsylvania test-and-packaging capacity and the acquisition of an NXP site in Arizona, the full-year range remains contained at €800–900 million.

Also worth noticing that management has said that component availability remains the principal industry constraint, which is one reason customers are placing longer-term orders. This is positive because it provides visibility, but it can also produce inventory accumulation, working-capital pressure and delayed cash conversion.

Nokia has already proved that demand exists. The next stage is proving that demand can be converted into profitable revenue and then into cash without requiring disproportionate inventory, capital expenditure or customer concessions.

Infinera Was the Bridge to the Data-Center Opportunity - Nokia’s AI infrastructure strategy.

Before the deal, Nokia already had a meaningful optical networking business.

Infinera added greater scale, stronger North American exposure, deeper relationships with webscale customers and complementary optical technologies. The combined business serves more than one thousand customers and has a broader portfolio spanning metro, long-haul, subsea, coherent optics and increasingly important high-speed pluggable products.

Infinera generated around 60 percent of its sales in the United States and more than 30 percent from webscale customers, with relatively limited customer overlap between the two companies.

That combination matters because the principal optical challenge created by artificial intelligence is not simply transmitting more data, network operators must increase capacity while reducing power consumption, physical footprint and cost per transmitted bit, they must also secure sufficient components in a supply chain that is already struggling to meet demand. So Infinera gave Nokia greater control over each of those variables.

The ICE7 coherent optical engine can transmit up to 1.2 terabits per second on a single wavelength, Nokia states that it can deliver 800G over distances exceeding 3,000 kilometers and support close to 100 terabits per second over a single fiber pair when Super C and Super L spectrum are used. This makes the technology applicable not only to short data-center connections, but also to metro, regional, long-haul and subsea networks connecting geographically distributed AI infrastructure.

Telefónica selected Nokia as the exclusive networking provider for 17 edge data-center nodes across Spain, with 12 already deployed. Aureon is using Nokia’s ICE7 and 1830 GX platform for a route that begins with 100 terabits per second of capacity and can scale to 400 terabits. Nokia, CSC and SURF demonstrated more than 1.2 terabits per second across a 3,500-kilometer European research network in preparation for the LUMI-AI supercomputer. Nokia has also partnered with Supermicro to integrate 800G switching hardware with SR Linux and Event-Driven Automation.

The less visible and potentially more important technology is ICE-D: Built on a monolithic indium phosphide photonic integrated circuit, ICE-D integrates lasers, modulators and photodetectors on a single optical chip. It can support up to 3.2 terabits per second and operate below 10 watts in certain linear-drive pluggable architectures. This moves Nokia beyond transporting data between data centers and gives it a potential position inside AI clusters, where power consumption and optical density are becoming fundamental constraints.

Nokia can also sell parts of this technology independently, its indium phosphide-based optical subassemblies can be integrated with either Nokia or third-party digital signal processors. This creates the possibility that Nokia participates in the economics of optical components even when another company supplies the final networking system.

The ramp is visible: in the second quarter of 2025, Optical Networks reported a book-to-bill ratio well above one, supported by new hyperscaler orders, and Nokia received its first award for 800G pluggables from a United States hyperscaler. By the second quarter of 2026, Nokia had booked €2.8 billion in AI and cloud orders, while Optical Networks sales grew 20 percent and sales to AI and cloud customers more than doubled.

The manufacturing strategy may be the most underappreciated part of the acquisition, as stated before, Nokia is ramping a new San Jose facility, expanding advanced test and packaging capacity in Pennsylvania by ten times and acquiring a semiconductor campus in Arizona that it intends to convert to indium phosphide production. These investments give Nokia greater control over supply at a time when optical component availability is limiting industry growth while they also create a domestically manufactured technology platform that may appeal to United States hyperscalers, government customers and critical-infrastructure operators.

This does not mean Nokia has an uncontested technological lead, competitors are all around. Ciena is developing 1600ZR and ZR+ pluggables, full-spectrum coherent transport and a 6.4T optical engine following its acquisition of Nubis. The optical market will remain intensely competitive, and headline transmission speed alone is not a durable moat.

But Nokia’s potential advantage is broader, it now combines DSP design, silicon photonics, indium phosphide, optical components, coherent engines, complete transport systems, IP routing, data-center switching and automation. That vertical breadth allows the company to optimize products for different combinations of reach, power, density and cost rather than relying on a single architecture.

The financial case is also becoming more credible. Nokia originally targeted €200 million of annual comparable operating-profit synergies by 2027 and more than 10 percent comparable EPS accretion from the acquisition.

In the second quarter of 2026, Network Infrastructure’s operating margin increased from 6.4 percent to 8.1 percent, while its gross margin rose from 40.3 percent to 42.7 percent. The company has not disclosed enough information to attribute those improvements entirely to Infinera, but the direction is consistent with the acquisition thesis.

I therefore now view Infinera less as a bridge to the data-center market and more as a platform acquisition. It gave Nokia technology, manufacturing capacity, webscale customer access and a credible route into both inter-data-center and intra-data-center optical networking.

If I had to summarize in just one sentence, Infinera was a superb move by Nokia’s management, it also made the Nvidia partnership possible. Not sure the market already fully recognizes that.

The NVIDIA Partnership

In October 2025, NVIDIA invested $1 billion in Nokia through a directed issuance of approximately 166.4 million new shares at $6.01 per share, equivalent to €5.16 at the exchange rate used for the transaction. NVIDIA became a 2.9 percent shareholder, while Nokia’s total share count increased to approximately 5.74 billion. The investment diluted existing shareholders by roughly 3 percent, but it also placed the world’s dominant AI computing company directly on Nokia’s shareholder register.

Infinera is helping Nokia demonstrate present-day growth through optical systems sold into data-center and cloud infrastructure. NVIDIA is changing the market’s assumptions about what Nokia might become.

The Partnership Has Two Separate Economic Engines

The first is the modernization of mobile radio infrastructure. Nokia is adapting its anyRAN software to NVIDIA’s CUDA ecosystem and accelerated computing platforms, with the objective of running 5G, 5G-Advanced, future 6G functions and selected AI workloads on shared infrastructure.

The second component is data-center networking. NVIDIA and Nokia agreed to work on data-center switching using Nokia’s SR Linux software with NVIDIA’s Spectrum-X Ethernet platform. They are also examining the use of Nokia’s telemetry and fabric-management technology inside NVIDIA AI infrastructure and exploring whether Nokia’s optical technologies could become part of future NVIDIA architectures.

Why “exploring”? We must be clear: NVIDIA has not committed to purchasing a specified volume of Nokia switches or optical equipment, and there is no public supply agreement guaranteeing that Infinera-derived technologies will enter NVIDIA reference architectures.

Nevertheless, the possibility is strategically important, if Nokia’s optical systems, coherent components, switching software or automation products become embedded in NVIDIA’s ecosystem, Nokia would gain access to AI infrastructure spending through a channel that reaches far beyond conventional telecom operators.

This is where the relationship with Infinera becomes especially relevant, Infinera gave Nokia greater scale in coherent optics, photonic integration, high-speed pluggables and North American webscale accounts. NVIDIA provides a potential route through which those capabilities could become more tightly connected to the architecture of AI factories.

Infinera therefore supplies the physical networking depth, while NVIDIA supplies ecosystem access and architectural validation. The two initiatives reinforce one another in a way that neither would achieve alone.

AI-RAN Has Moved Beyond the Presentation Stage

In July 2026, Nokia announced what it described as the industry’s first commercial AI-native RAN platform.

It is based on Nokia’s anyRAN software and NVIDIA’s accelerated computing, with pilot deployments scheduled for the end of 2026 and commercial availability planned for 2027.

The platform provides operators with three adoption paths.

• Existing Nokia customers can install a GPU-powered capacity card inside deployed AirScale baseband systems. This reduces the need for a complete equipment replacement and gives Nokia a potentially large installed-base upgrade opportunity.

• Operators requiring additional capacity can deploy a standalone GPU-powered AI-RAN node supporting 4G, 5G and eventually 6G workloads.

• Customers pursuing cloud-native networks can run Nokia’s AI-RAN software on commercial servers supplied through an ecosystem that includes Dell, Supermicro, Quanta and Red Hat.

The diagram illustrates how Nokia’s AI-native RAN architecture can be deployed through three migration paths: upgrading existing AirScale systems, adding dedicated GPU-powered nodes, or using cloud-native infrastructure. The architecture supports radio processing, AI inference, network optimization, and future 5G-Advanced and 6G capabilities. It also enables operators to modernize networks gradually rather than replacing entire systems at once. Economically, the model could shift part of Nokia’s RAN business from irregular hardware sales toward recurring software and subscription revenue, although pricing, margins, and commercial adoption remain unproven.

Operators do not have to rebuild an entire network around NVIDIA infrastructure immediately, they can add accelerated computing to existing Nokia systems, deploy it selectively in high-traffic areas or introduce it through cloud-based RAN infrastructure.

That creates a more credible commercialization path than requiring a synchronized industry-wide transition to an entirely new architecture, a process telecom operators would probably complete around the same time that glaciers finish their own strategic transformations.

Under the proposed structure, operators would pay for continuous access to AI algorithms, spectral-efficiency improvements, optimization functions and future network capabilities. Performance enhancements could be delivered through software rather than waiting for the next major hardware replacement cycle.

This potentially changes the economic character of part of Nokia’s Radio Networks business because traditional RAN is dominated by large, irregular equipment purchases. Revenue depends on operator capital-expenditure cycles, while pricing is pressured by a small number of powerful customers.

Hardware generations eventually become obsolete, forcing vendors to invest continuously in new custom silicon and product development.

A subscription model could introduce more recurring revenue, higher software content and better visibility. It could also allow Nokia to monetize performance improvements throughout the life of the installed equipment rather than capturing most of the value at the initial sale.

No pricing, contract duration, gross-margin profile or expected subscription revenue has been disclosed, that’s important. Still, this business-model evolution is one reason the market may be willing to assign Nokia a higher multiple before AI-RAN produces substantial revenue. The re-rating is therefore not based only on the number of AI-RAN systems Nokia might sell, but it also reflects the possibility that AI-RAN changes how Nokia earns money from its installed base.

The Technical Proposition Is Becoming More Concrete

Nokia says its AI-RAN platform has already demonstrated spectral-efficiency gains of more than 20 percent and is now targeting improvements of approximately 50 percent by 2027 and more than 100 percent by 2028, potentially allowing operators to carry twice as much traffic over the same spectrum. These remain company targets rather than independently verified results from large commercial networks.

If achieved, the economic implications could be substantial. Spectrum licenses are among the most expensive assets owned by mobile operators. Extracting more capacity from existing spectrum may be more attractive than purchasing additional frequencies or installing a large number of new sites.

The architecture could also reduce the frequency of silicon replacement. Nokia’s traditional baseband systems have depended heavily on purpose-built chips, with major hardware generations commonly changing every three to five years. NVIDIA’s programmable architecture is intended to support new functions and eventually 6G through software upgrades.

Nokia executives have claimed that the new GPU-based products can operate within a power and cost envelope comparable to the company’s custom-silicon products. They have also argued that an NVIDIA-based platform could remain useful deeper into the 6G cycle than current purpose-built hardware. These assertions remain to be proven under commercial network conditions, but they directly address two of the strongest criticisms of AI-RAN: GPU cost and power consumption.

AI-RAN is designed to allow RAN and external AI applications to share the same computing resources. During periods when radio workloads use less capacity, spare GPU resources might support inference, video analytics, robotics, industrial applications or other edge AI services.

Instead of operating base stations solely as a cost center, telecom providers could theoretically sell distributed computing services from network locations close to users and machines.

A 2026 academic techno-economic study found that, under certain assumptions, revenues from external AI workloads could offset the additional costs of GPU-heavy RAN deployments and generate substantial returns. The authors also emphasized that the economic case remains highly dependent on demand, utilization, token pricing and deployment density.

How would it work commercially?

NVIDIA would sell accelerated computing platforms, expand CUDA into another large industry and influence future network standards. Nokia gives it commercial RAN software, operator relationships, radio expertise and a practical path into live networks.

Nokia receives capital, access to the CUDA developer ecosystem and a merchant-silicon roadmap funded by a company spending vastly more on accelerated computing than Nokia could justify independently.

The partnership may also reduce some of Nokia’s internal silicon-development burden. Nokia has indicated that it is making a significant transition from custom baseband silicon toward merchant accelerated computing, although it continues to use Marvell technology in parts of the portfolio and is attempting to preserve hardware abstraction within its software.

This creates both leverage and dependence.

Nokia gains access to NVIDIA’s performance roadmap, but it may also become more exposed to NVIDIA’s pricing, product availability and proprietary CUDA ecosystem. Ericsson has criticized this approach as a possible source of vendor lock-in and continues to emphasize custom silicon and more hardware-neutral virtual RAN software.

To summarize, the partnership began with investment and architectural intent in October 2025. It moved to functional validation and multiple operator integrations by March 2026. Nokia introduced a defined commercial platform and business model in July 2026. Pilot deployments are expected near the end of the year, followed by commercial availability in 2027.

NVIDIA and Infinera Are Re-Rating Nokia Together - What’s next to evaluate?

It should be clear by now that both Infinera and NVIDIA affect Nokia’s valuation through different time horizons.

Infinera is the near-term earnings engine. It strengthens Nokia’s optical portfolio, expands its webscale customer relationships and gives the company greater exposure to current AI data-center construction. Optical and IP growth, AI and cloud orders and improving Network Infrastructure margins provide observable financial evidence.

NVIDIA is the duration engine. It increases the possible length and breadth of Nokia’s growth runway by connecting the company to AI-RAN, edge computing, data-center switching, software subscriptions and 6G.

So the market has to assign value to the following several options:

• Nokia could become a larger supplier of optical and switching technology to AI infrastructure.

• Its SR Linux software could gain relevance through NVIDIA’s Spectrum-X ecosystem.

• Its installed AirScale base could become a distribution channel for AI-accelerated upgrades and software subscriptions.

• Its radio sites could eventually support external edge-AI workloads.

• Its anyRAN software could become an important layer in AI-native 6G networks.

Lets discuss them:

1. Nokia becomes a larger supplier of optical and switching technology to AI infrastructure. Estimated probability: 92 percent

This is the most advanced part of the thesis and should already be considered part of the base case.

In the second quarter of 2026, Nokia’s sales to AI and cloud customers increased 105 percent, while Optical Networks grew 20 percent and IP Networks grew 16 percent. The company received €2.8 billion in AI and cloud orders and expects approximately half of that amount to convert into revenue over the following twelve months.

Commercial evidence extends beyond order intake. Nscale has made Nokia a preferred networking partner for its global AI infrastructure expansion and already uses Nokia’s data-center switching and routing technology at its Stavanger facility. Telefónica selected Nokia as the exclusive networking provider for 17 edge data-center nodes in Spain, of which 12 have already been deployed. Nokia’s 800G coherent pluggables are shipping to a large United States customer, while Aureon is using Nokia’s ICE7 optical technology in a network capable of scaling to 400 terabits per second.

2. SR Linux gains relevance through NVIDIA’s Spectrum-X ecosystem. Estimated probability: 50 percent

NVIDIA and Nokia have formally agreed to collaborate on data-center switching using Nokia’s SR Linux software with the Spectrum-X Ethernet platform. They are also evaluating Nokia’s telemetry and fabric-management technology for NVIDIA AI infrastructure and exploring the possible inclusion of Nokia optical technology in future NVIDIA architectures.

The integration has progressed beyond a memorandum. SR Linux is now represented in NVIDIA DSX Air, allowing AI cloud builders to simulate, validate and automate Nokia-based network environments before physical deployment.

What remains absent is a publicly disclosed production customer using SR Linux on Spectrum-X or material revenue directly attributable to the integration.

SR Linux clearly has independent product relevance, as demonstrated by deployments with Nscale and partnerships with Supermicro. The uncertainty concerns how much additional distribution NVIDIA will provide and how much of the resulting economics Nokia can retain.

3. Nokia’s installed AirScale base becomes a distribution channel for accelerated upgrades and subscriptions. Estimated probability: 78 percent

Nokia has introduced a GPU-powered capacity plug-in designed specifically for existing AirScale baseband systems. Operators can add accelerated computing without replacing the entire chassis or radio infrastructure, the company has also announced a subscription model through which customers would receive continuing access to AI algorithms, spectral-efficiency enhancements and network-optimization capabilities. Pilots are expected by the end of 2026, followed by commercial availability in 2027.

This creates a credible channel for monetizing Nokia’s installed base twice: first through incremental computing hardware and then through recurring software.

Tests with T-Mobile and Indosat have already demonstrated Nokia RAN software operating on NVIDIA-accelerated infrastructure in operator environments. BT, Elisa, NTT DOCOMO and Vodafone are also involved in development or evaluation.

The remaining uncertainty is commercial rather than architectural. Nokia has not disclosed subscription prices, contract durations, customer volumes or expected margins. Operators also remain cautious about GPU cost, energy requirements and dependence on the CUDA ecosystem.

4. Radio sites support external edge-AI workloads. Estimated probability: 35 percent

Nokia and SoftBank have demonstrated that spare AI-RAN computing capacity can be identified and allocated to third-party AI tasks. T-Mobile has separately demonstrated concurrent RAN processing and AI applications on a single NVIDIA Grace Hopper server using live spectrum and commercial radio equipment.

For the model to work, operators must find customers requiring low-latency local inference, maintain sufficiently high GPU utilization and compete economically with centralized cloud providers. They must also operate distributed computing infrastructure across locations originally designed primarily for telecommunications equipment.

External edge workloads may eventually improve the return on AI-RAN investment, but I would assign only modest value to this option until operators disclose paying customers, utilization rates and pricing.

5. anyRAN becomes an important software layer in AI-native 6G networks. Estimated probability: 60 percent

Nokia’s anyRAN software has already been validated across NVIDIA-accelerated infrastructure and is being evaluated by a growing operator and technology ecosystem. The commercial AI-RAN platform announced in July 2026 uses the same software foundation across upgraded AirScale systems, dedicated AI-RAN nodes and cloud-native server deployments, with an intended software path from 5G and 5G-Advanced to 6G.

The broader direction is consistent with early 6G standardization. The ITU has included Artificial Intelligence and Communication as an official IMT-2030 use scenario and identified ubiquitous intelligence as a design principle. Final radio-interface standards, however, are not expected until the end of the decade.

Nokia therefore has an early architectural position, not a guaranteed standard. Operators may also demand hardware neutrality that limits the role of any NVIDIA-centered architecture.

Conclusion.

The probability that at least three of the five options become financially relevant by 2030 is, in my view, approximately 70 percent. The probability that every option succeeds materially is closer to 10 or 15 percent.

This is the central discipline required when valuing Nokia. The market is justified in recognizing that the company now owns several credible paths into AI infrastructure. It is not justified in valuing every path as though it has already reached commercial scale.

Competition: Stronger in Optics, Still Challenged in Radio and Data Centers

Nokia competes across several distinct infrastructure markets, which makes its competitive position more complex than a simple comparison with Ericsson or Huawei.

In mobile radio access networks, Ericsson remains its most direct Western rival, while Huawei continues to compete aggressively in markets where regulatory restrictions do not limit its presence. Samsung has also gained relevance in selected geographies, and emerging Open RAN architectures create additional pressure from software vendors and semiconductor companies. Nokia’s AirScale portfolio has become more competitive, and the NVIDIA partnership gives it a differentiated AI-RAN roadmap, but traditional RAN remains a cyclical, price-sensitive market with relatively modest margins.

Following the acquisition of Infinera, it now has a broader exposure to North American and webscale customers, and deeper capabilities in coherent optics, pluggables and photonic integration. Its principal competitors include Ciena, Huawei, Cisco through Acacia, ZTE and NEC. Nokia is now one of the largest global optical suppliers, but the market remains technologically demanding and highly competitive. Its advantage is increasingly based on the breadth of its portfolio, combining optical transport, IP routing and network automation rather than relying on a single product category.

In IP routing and data-center switching, Nokia faces a more difficult competitive landscape. Cisco, Arista and Juniper have deeply established customer relationships, large installed bases and strong positions in cloud and enterprise networking. Nokia’s 7750 and 7220 platforms, together with SR Linux, provide credible technology, and recent partnerships with NVIDIA, Supermicro and AI infrastructure operators could improve its relevance.

Core network software and automation offer a more balanced opportunity. Nokia competes with Ericsson, Huawei, Oracle, Cisco and other cloud-oriented vendors, but its existing operator relationships and experience managing complex telecom networks provide an important advantage. As network functions move from specialized hardware toward cloud-native software, automation and orchestration may become more valuable sources of recurring revenue. Nokia’s challenge is to increase software margins while avoiding commoditization as operators demand greater interoperability and vendor flexibility.

Technology licensing remains one of Nokia’s strongest competitive assets. Its extensive portfolio of cellular patents provides high-margin recurring revenue and places the company alongside Qualcomm, Ericsson and Huawei as one of the major owners of essential communications intellectual property. This business supports research and cash generation, although revenue can fluctuate because of contract renewals, litigation and catch-up payments.

Concluding, Nokia has to defend its telecom franchise, convert its optical scale into higher margins and use partnerships such as NVIDIA to gain relevance in markets where Cisco, Arista, Ciena and Ericsson currently hold stronger positions.

The Real Bet for Nokia

At the end of this article I hope my readers have a much broader view on Nokia current challenges and landscapes, much further than simply “optics orders are growing QoQ” or “Nokia has partnered with NVIDIA.”

The investment thesis therefore does not depend on Nokia becoming an “AI company.” That label is too vague to be useful for Nokia.

Infinera makes Nokia more relevant to AI infrastructure today. NVIDIA makes its longer-term software and mobile strategy more credible. The installed telecom base gives those technologies a path into commercial networks that a semiconductor company could not reproduce quickly on its own.

If its optical technology becomes embedded in a growing number of AI networks, SR Linux earns a role inside accelerated data-center fabrics, and anyRAN turns part of the AirScale installed base into programmable infrastructure, Nokia will have created several reinforcing distribution channels around the same underlying demand. A customer acquired through optical networking could adopt its routing and automation. A mobile operator already using AirScale could purchase accelerated upgrades and software. Research funded by licensing income and Bell Labs could strengthen products across both infrastructure segments.

If Nokia succeeds, the most important result will not be a single quarter of higher optical growth or a successful AI-RAN pilot.

It will be that, almost unnoticed, a company once defined by telecom equipment became part of the infrastructure through which artificial intelligence actually moves.

Today, the company combines a global distribution business with Hyve Solutions, a design and manufacturing unit for data center, cloud, and accelerated computing infrastructure, alongside a growing layer of software, cloud commerce, services, and enablement. This combination became more apparent following the 2026 segment reorganization, when Hyve began to be reported separately from the regional distribution businesses.

It isn’t a pure AI play, but no problem, because the recent acquisitions increase the likelihood of a structurally better mix, especially because they add proprietary software, cloud orchestration, migration services, and cross-selling capabilities in key markets. The acquisition of Apptium is the most strategic move: it brings key technology from the StreamOne platform in-house, accelerates the XaaS/SaaS/IaaS/agentic AI ecosystem, and boosts monetization across software, billing, automation, and marketplace syndication, business lines that typically offer better margins than pure hardware resale. IPsense adds AWS migration services in Brazil and Latin America, while Hyve originated within legacy SYNNEX and gained access to a much broader commercial platform.

Today has all that plus a growing legacy business, but still trades at a low multiple of 18 P/E, with no premium at all. Will it change?

The most important data point for the margin thesis is that Hyve already operates with an operating margin well above that of the distribution business: in 2Q26, Hyve’s operating margin was 6.03%, versus 2.55% in the Americas, 1.19% in Europe, and 2.47% in APJ for distribution. Over the six-month period, Hyve reached 6.53%, compared to 2.54%, 1.47%, and 2.72%, respectively.

At approximately US$250.16 per share, with a market cap of US$19.95 billion and a current P/E ratio of 17.9x, using FY25 metrics, this equates to approximately 12.4x EV/Adjusted EBITDA and a 7.1% FCF yield. As of July 2026, recent post-Q2 targets were raised by Morgan Stanley (US$374), UBS (US$352), RBC (US$340), Goldman (US$300), Barrington (US$325), and Barclays (US$278, Equal Weight).

How TD SYNNEX’s Evolving Business Mix Could Support Margin Expansion and a Re-Rating

Read more

To understand where this value hides, picture what actually happens inside a machine that carves circuits into a silicon wafer, that carving is done with plasma, which is just gas energized until it becomes a violent soup of charged particles.

The wafer sits in the chamber while the plasma rains charged particles down onto it to etch the pattern, and right around the rim of the wafer sits a ring of material whose only job is to keep that rain falling evenly all the way out to the edge, just to make sure the pattern at the edge is maintained. That ring is called a focus ring, and here is the catch: the same plasma the ring is shaping also eats the ring, it cannot do its job without standing in the line of fire, so it is destroyed on purpose, a little more with every wafer that passes through, pulled out and replaced, then either scrapped or sent into a cleaning/recoating loop. This sentence is the entire business in miniature.

A part that is destroyed on a schedule is a part that has to be bought again on a schedule, which turns a lump of ceramic into something far more valuable: a recurring revenue stream agnostic to all chipmakers running these chambers that burns through the same rings.

The company that has owned this particular niche is a Korean one almost nobody outside the industry has heard of, Tokai Carbon Korea, listed in Seoul under the code 064760, which by its parent's own account once held around 80 percent of the world market for the solid silicon-carbide version of these rings.

The rim of the wafer actually gets hit a few times harder than the ring does, which sounds like the ring gets off easy, except that even the lighter dose wears it down to replacement inside two weeks. Tokai Carbon Korea sells silicon carbide focus rings, the most difficult and patented component to manufacture, with very few competitors capable of producing them to the required quality.

Why the material keeps changing, and who that pays

There is a second way these parts die, in the furnaces the enemy is not plasma but heat and time, because the trays, tubes and liners that hold the wafers are cycled to extreme temperatures over and over, and the traditional material, quartz, is wonderfully pure but mechanically weak, so it sags and warps under that repeated stress and has to be replaced. This is why the industry has spent decades slowly swapping structural quartz for silicon carbide, a far tougher ceramic, in the parts that have to hold their shape.

Since those parts are also potential wafer contaminants in disguise, purity has a big role in this layer. A trace of metal hiding inside the material can creep out under high heat, land on the wafer, and ruin it, so making one that holds its impurities prisoner rather than leaking them is genuinely hard, so definitely not a commodity.

So the layer is really a competition between three families of material, each winning a different corner of the chamber, and each corner has its own set of suppliers.

Quartz keeps winning in the gentler furnace zones where purity matters most, which is the territory of the quartzware makers, and the cleanest listed way to own that corner is in Korea again, through Wonik QnC, code 074600, which is no minor name but the clear Korean leader in chamber quartzware, holding something close to a third of the global market by some estimates, running an operating margin near the mid-teens, and bought the quartz business of the old American supplier Momentive to extend its reach into the United States and Europe.

Wonik QnC sells quartzware, which includes quartz tubes, boats, and liners. These are critical components requiring purity, but they are technically less exotic than SiC rings, and there are more manufacturers capable of producing them worldwide (Shin-Etsu, Tosoh, Heraeus, and emerging Chinese companies). More competition means less power to dictate prices, and this compresses margins.

Tellingly for this thesis, it also runs its own cleaning-and-coating arm, the regeneration business, which grew more than twenty percent in 2025, so it straddles both halves of the value chain at once. The classic quartz names a Western investor would reach for first, Japan's Shin-Etsu Quartz and Tosoh Quartz, the German group Heraeus and the privately held Momentive, are all industrially serious but give you no clean way in, because usually they are joint ventures, subsidiaries or private.

Silicon carbide wins the second corner, the violent plasma chambers, in its highest-purity form grown atom by atom in a process the makers call CVD, because it shrugs off the plasma, heat and corrosion that quartz and plain silicon cannot survive. Tokai Carbon, the Japanese parent of that Korean champion, describes the migration of focus rings from silicon and quartz toward silicon carbide explicitly, and ties it directly to the rising number of etch steps and the climbing plasma power that came with the move to three-dimensional chips, which is the single most important sentence in the entire thesis, because it means the material shift is a forced response to the physics getting harder. Silicon carbide became the premium material at exactly the moment the process became too dense, too hot, and too expensive for mistakes.

The third corner, the coatings on the chamber's inner walls, belongs to a ceramic called yttria, which has been steadily replacing an older standard called alumina because it survives the corrosive gases better and sheds fewer of the stray particles that land on wafers and wreck them, but the coatings are a smaller prize than the rings and tend to live inside the bigger players rather than in pure plays of their own.

What really matters

It is that each advanced wafer consumes more of the chamber around it, so the right way to think about demand is consumption per wafer, not wafers per month. Tokai Carbon was clear on the mechanism, that the move to three-dimensional chips raised both the number of etch steps and the plasma power, and that this accelerated demand for its solid silicon-carbide rings even before you count a single extra wafer.

The silicon-carbide ring lasts a little longer than the old silicon one, it costs far more, with Korean research putting the price of a silicon-carbide ring around two and a half times that of a silicon one. The migration to the premium material therefore lifts spending per chamber even if nothing speeds up the replacement clock, which means the right supplier is not just riding volume growth, but it is riding a physics upgrade, the slow ratchet of the chamber getting more violent and the parts inside it getting more expensive to keep replacing.

This is also where the second serious Korean name enters, because the company best placed to ride that volume without paying for Tokai Carbon Korea's pedigree is Hana Materials, code 166090, which makes silicon electrodes and silicon rings with a growing silicon-carbide line, and is essentially a pure bet on etch consumables, with 2025 revenue around 273 billion won and a market value near 1.3 trillion won.

Hana plays in a less protected part of the chamber than the focus ring monopoly, which sounds like a weakness and is actually the source of its appeal, because it is the name most likely to pick up incremental share if the big memory makers push to add a second supplier, a push that is exactly what the next section is about.

There is an even smaller and more obscure version of the same idea in Worldex Industry & Trading, code 101160, a Korean maker of silicon electrodes, rings and boats that sends something like forty percent of its output to Samsung and SK Hynix.

Subscribe now

The monopoly under attack

Now back to that Korean champion, Tokai Carbon Korea, because its story is my favorite kind, more interesting than a simple monopoly, since three things are true about it at once.

The first is that it really did dominate, with its parent claiming roughly 80 percent of the world market for solid silicon-carbide focus rings and Korean trade press long describing it as controlling more than 80 percent of the Korean market, which is about as close to a monopoly as this industry produces. The second is an honesty caveat worth stating plainly, and it bugged me more than I expected, that I simply could not find an officially updated share figure for 2025 or 2026, so the 80 percent should be read as a strong historical fact rather than a confirmed snapshot of today. The third, and the reason the stock is a live debate rather than a sleepy compounder, is that the patent wall protecting all of this has come down by half.

The legal turn came in March 2023: the Korean Supreme Court let stand a ruling that the company's patent on the properties of the silicon-carbide ring was invalid, while its patent on the manufacturing method survived. So the wall fell halfway. A challenger no longer has to accept that the ring's properties are off limits, but it still has to find its own way around a real protection on how the ring is actually made, which is precisely why challengers appeared and yet did not storm the market overnight. The incumbent then went on the legal offensive against other names including Hana Materials and a small challenger called KNJ.

The margin picture needs the same honesty as the market-share one, because the comfortable shorthand that this is a mid-thirties-percent operating margin business was true in the boom but is not the current photograph. The company reported 2025 revenue of about 301 billion won and operating profit of about 84 billion won, which is a margin near 28 percent, down from nearly 40 percent back in 2023, so the accurate sentence is that this is a structurally very profitable business whose margin has cooled from spectacular to merely excellent. Still remarkable for a parts maker, and it still tells you how much pricing power lives in a component that sits in the line of fire and that the customer has slowly, expensively, learned to trust.

The challenger thesis has a sharpest expression, and its name is KNJ, code 272110. It is the smallest and most speculative of the Korean names, with 2025 revenue around 84 billion won, and it claims to have been the second company in the world to mass-produce the CVD silicon-carbide focus ring

The afterlife of a sacrificed part

There is a second economic layer sitting on top of the first. A consumable that gets eaten does not simply get thrown away, so you have a cleaning-and-coating cycle, which recurring business is stretching the life of an expensive component and holding yield steady.

Long before the part is fully spent it gets pulled, shipped out, stripped of its contaminated and eroded surface, recoated, requalified, and sent back into the chamber to be sacrificed all over again. The global market for this refurbishment work runs somewhere around 1.6 to 2.2 billion dollars and grows as the processes get harder.

The leader of that business is another Korean name hiding in plain sight, KoMiCo, code 183300, which pioneered the cleaning-and-coating service in Korea back in 1996 and is now the global leader in it, and whose numbers show how real the AI buildout made this cycle, with revenue jumping roughly 65 percent in 2024 to around 507 billion won, ₩604B in 2025 with the stock more than doubling over the following year as fabs ran their chambers harder and dirtier. KoMiCo has even been following its customers across the Pacific, picking up a Texas incentive in 2025 to expand near the new American fabs, which is the same supply-chain relocation pulling the whole Asian consumables ecosystem toward Arizona and Texas, and the Samsung-facing alternative in the same business is Hansol IONES, code 114810, part of the Hansol group.

For an investor the useful conclusion is that the recurring revenue in this theme runs in two streams rather than one, the making of the part and the resurrection of it, and the genuinely elegant thing, the part I kept coming back to, is that they hedge each other, because furious new-part consumption favors Tokai Carbon Korea while furious regeneration favors KoMiCo.

Why the alpha is Korean, not Japanese

The pattern you have probably noticed by now is the most useful, and frankly the most surprising, finding of this whole search, which is that almost every clean way to own this theme is Korean, even though the instinct of most investors hunting for semiconductor consumables, mine included when I started, is to look in Japan.

The reason is structural rather than accidental, because in Japan the best parts of this business are buried inside large diversified groups, while in Korea they trade as focused, listed companies you can actually buy.

Toyo Tanso, listed as 5310, is a high-quality maker of specialty graphite and silicon-carbide-coated graphite that expects about half its sales to come from semiconductors, but it is more a bet on the silicon-carbide power-chip cycle than a pure ring play.

Ferrotec, listed as 6890, is a sprawling picks-and-shovels platform that makes quartz, CVD silicon carbide, silicon parts and cleaning services all at once, but its group operating margin near 9 percent tells you it is a diversified industrial rather than a focused compounder.

And Tokai Carbon itself, the parent listed as 5301, is the purest illustration of the trap, because its fine-carbon division is genuinely excellent and sells mostly into semiconductors, yet that jewel is, sadly, buried inside a conglomerate whose other divisions are more cyclical and less interesting, so you cannot buy the good business without buying the rest.

Another one is GLTECHNO Holdings, listed as 255A, the holding that absorbed a previously listed quartz-jig maker, which lets you own listed Japanese quartzware without paying for a giant, at the cost of a holding-company discount.

If I had to compress the whole thing into one sentence: the hidden money sits less in the machine and more in the material that takes a beating inside it. The parts in the chamber are built to be sacrificed, wafer after wafer, so someone has to keep making them and someone else has to keep bringing them back, to a tolerance almost no one can hit. That is a very short list on both sides and it is mostly Korean, not Japanese.

Set against that, the Korean names are what a focused investor actually wants, because Tokai Carbon Korea is the dominant ring maker, Hana Materials and the smaller Worldex are the optionality on a second source, KNJ is the asymmetric challenger, Wonik QnC is the quartz-and-ceramics play that also regenerates parts, and KoMiCo is the regeneration leader, and every one of them trades as its own company rather than as a line item inside something larger.

Where the thesis might break

The honest case against this trade starts with the obvious risk, that the incumbent's moat can be chipped without being destroyed, because the court already did half the challengers' work by invalidating the properties patent, so if the big memory makers accept more aggressive dual-sourcing, part of the premium leaks away.

The counter-argument is that the manufacturing-method patent still stands and that getting a brand-new supplier's part approved for a process this unforgiving is genuinely slow and hard, so the realistic picture is a premium narrowing rather than collapsing, which is bad for the monopoly story and good for the challengers, exactly the trade Hana and KNJ represent.

The second risk is the cycle, this is a better pocket than raw equipment, but it is not immune to memory's brutal swings, and if capex stalls, the recurring nature of the consumable softens the fall without preventing it.

The third risk is that the material war never stops, since silicon carbide's win over quartz and silicon is real today but the next process generation could favor yet another material, which is a reminder that a lead built on a specific material is always rented, never owned.

For the focus ring itself, silicon carbide has no clear successor in sight, since the ceramic alternatives lose to it on conductivity or toughness and the lab-stage candidates are still early, so the ring is more exposed to competition on price than to replacement.

The fourth is China, because both Tokai Carbon and Ferrotec are building capacity there, which helps cost today but opens a geopolitical flank tomorrow, and the deeper threat is not export controls but Chinese rivals learning to make these parts faster than anyone would like.

You might also like

If you enjoyed reading or you are new here, I highly recommend my post about the Timing Layer, another free and simple post explaining another overlooked and essential sector in AI infrastructure.

The global automation industry is undergoing a structural transition that is more profound than most market commentary acknowledges. The shift is not simply from manual to automated production, but from rigid, fixed-program mechatronic systems to adaptive robotic platforms governed by physical AI, machines that are able to perceive, simulate, and learn rather than merely execute predefined sequences. Japan sits at the center of this transition, not because it dominates the software layer, but because it controls the mechanical and sensory infrastructure without which no robotic system can function at scale.

So here we are going to map that ecosystem honestly: where the real moats are, which companies are genuinely advancing in physical AI versus those riding the narrative, where the revenue and profit actually concentrate, which partnerships signal structural positioning rather than marketing, and where the red flags are that most coverage glosses over.

Read more

The Real Moat: Tacit Manufacturing Knowledge

Before going through individual companies, it is worth understanding why the concentration described in this article is so difficult to dissolve even with substantial capital and political will. The standard framing for supply chain risk treats the problem as one of intellectual property, scale, or geography. Those matter, but they are not the deepest barrier. The deepest barrier is tacit process knowledge, and it is almost impossible to transfer quickly.

Tacit knowledge in semiconductor manufacturing is the accumulated understanding of why a process works, not just the documented description of what the process does. A T-Glass specification sheet describes the target dielectric constant, the fiber diameter, and the weave pattern. It does not describe the specific furnace atmosphere adjustments that prevent defect formation during the glass draw, the maintenance schedules that keep fiber diameter variance within tolerance, or the decades of trial and error that went into finding the glass composition that achieves the required electrical properties while remaining manufacturable at commercial yields. That knowledge lives in the people, procedures, and institutional memory of companies like Nittobo, and it cannot be extracted by reading patents or hiring a few engineers away.

The same logic applies across the supply chain. CMP slurry formulations describe abrasive particle size and chemistry. They do not describe the supplier-fab co-development loops that determined which slurry behavior under edge conditions correlates with yield loss on a specific device layer at a specific node. EUV resist specifications describe sensitivity and resolution. They do not describe the coating uniformity adjustments that suppress line edge roughness under specific scanner illumination conditions at a given fab. Quartz oscillator datasheets describe phase noise floors. They do not describe the crystal growth procedures that minimize defect density in the way that produces stable aging characteristics over a twenty-year product lifetime.

This is why “China can manufacture it” rarely means “China can replace it” in this segment of the supply chain. A Chinese company can build a factory, purchase equipment, and hire chemists. It can produce a material that meets the written specification. What it cannot quickly acquire is the process stability and yield consistency that incumbent Japanese suppliers have built through decades of iterative refinement with their fab customers. Fabs do not simply test whether a material meets specification in isolation. They qualify it across their full process flow, across temperature and humidity ranges, across multiple equipment platforms, and across the variation in incoming raw material batches that occurs over years of production. That qualification captures tacit knowledge that is not in any document, and it is why a new entrant with a technically adequate product can spend two to three years in qualification before receiving production approval.

The implication for investors and policymakers is that the substitution timelines described throughout this article are not primarily engineering constraints or capital constraints. They are knowledge transfer constraints, and those are slower and more expensive to overcome than any other kind.

I also want to highlight that chokepoint does not always equals higher margins. Japanese business culture is not towards to EPS but sustainbability and community driven. They may be relutant to increase prices at will, as we have seen happening recently with Nittobo and others.

Lasertec: The One Company Standing Between EUV and Production

If there is a single Japanese company whose absence would most immediately stop advanced semiconductor production globally, it is Lasertec. The company produces actinic mask inspection systems, which are the tools used to check EUV photomasks for defects before those masks are used to pattern chips at the 7nm node and below. ASML makes the EUV scanners that perform the actual lithography. But before a mask can go into an ASML scanner, it needs to pass inspection, and Lasertec is the only company in the world that makes the equipment capable of performing that inspection using EUV-wavelength illumination, which is the only method accurate enough to detect defects at the feature sizes involved.

There are no qualified alternatives for actinic inspection specifically. Competing inspection approaches using deep ultraviolet illumination exist and are used for less demanding mask layers, but they cannot reliably detect the defect types that matter at advanced EUV nodes. KLA Tencor dominates the broader wafer and reticle inspection market and competes in certain e-beam mask inspection categories, but it does not have a product that replaces Lasertec’s actinic systems for EUV mask qualification at leading nodes. The competitive moat here is not primarily intellectual property or scale, though both matter. It is that developing actinic inspection capability requires building a coherent EUV light source, integrating it with precision optics and vibration isolation, and writing software capable of distinguishing printable defects from measurement noise at sub-nanometer feature sizes. That is a program that takes a decade and hundreds of millions of dollars, and no company outside Japan has prioritized it.

TSMC, Samsung, and Intel all depend on Lasertec systems to qualify their EUV masks. Every advanced node chip produced today, meaning every chip at 7nm and below including AI accelerators, high-end CPUs, and leading-edge memory, passed through a Lasertec inspection system before it was manufactured. A prolonged disruption at Lasertec would effectively halt new EUV mask qualification and severely constrain production ramps, since fabs maintain some inventory of qualified masks but cannot introduce new designs or respond to mask defects without inspection capability.

The substitution timeline for Lasertec is not measured in months or even years. It is measured in the time required to build a credible actinic inspection program from scratch, which no government or company has announced doing in a serious way. The geopolitical risk here is asymmetric in a way that most semiconductor supply chain analysis has not fully absorbed.

Hitachi occupies a related but distinct position in the metrology and inspection layer. The company produces critical dimension scanning electron microscopes and electron beam defect review systems used to measure feature dimensions and characterize defects after each patterning step. Where Lasertec inspects the mask before it enters the scanner, Hitachi High-Tech measures what the scanner produced on the wafer afterward. CD-SEM measurement is required at every critical layer of an advanced process flow, and Hitachi’s installed base at leading fabs is dominant enough that a supply disruption would affect process control at TSMC, Samsung, and every other advanced node manufacturer simultaneously. KLA competes in certain inspection and review categories, and Applied Materials has metrology products, but Hitachi’s electron optics heritage and decades of co-development with fab process engineering teams give it a position in CD-SEM that is difficult to displace regardless of competitive investment.

HOYA produces the EUV mask blanks that sit one step upstream of Lasertec in the same production sequence. A mask blank is the ultra-flat synthetic quartz substrate onto which the chip pattern is etched before the mask goes into the Lasertec inspection system and then into the ASML scanner. Without a qualified blank, there is no mask to inspect and no mask to expose. HOYA holds the dominant global position in EUV mask blanks, with AGC as the only other supplier operating at meaningful scale. The structural difference between HOYA and Lasertec is that blanks are a consumable, each one is used once and discarded after the mask reaches end of life, which means demand scales directly with EUV production volume rather than with capital equipment cycles. As EUV layer counts per chip increase with each process node, blank consumption per wafer starts increases proportionally, and HOYA’s position tightens with every generation rather than stabilizing.

Read more

The House of Cards of 1.6T

In the transitions from 400G to 800G, there was engineering headroom. Components could underperform at the margins, the system would still close the link, and the penalties were manageable but at 1.6T, each lane runs at 224G using PAM4 modulation, which compresses the unit interval, a time window available to transmit each bit, to a few picoseconds. That compression does not affect one subsystem, but all of them simultaneously, and it does so through a shared constraint.

The error budget is the total amount of signal degradation a system can tolerate while still recovering data reliably. At lower speeds, that budget was large enough that jitter, thermal drift, PCB insertion loss, connector loss, DSP noise, laser instability, and crosstalk could each be treated as separate engineering problems with independent tolerances. At 1.6T, they are no longer independent, it is so fast that every imperfection competes for the same residual margin, and the margin is small enough that the competition is real. An oscillator that performs adequately in isolation may behave differently once the thermal profile of the switch changes, and that change may push the system over threshold if the PCB already consumed most of its insertion loss budget and the laser is running at the edge of its linearity.

This is the structural shift that separates 1.6T from prior generations. Every component is no longer evaluated on whether it works, but on how much margin it consumes, and in a system with almost no margin left, that is a different and more demanding standard, and that is where we are heading to.

The supply chain around 1.6T looks structurally different from the supply chain around 800G for this reason. The physics imposes a higher tier of analog precision on every layer of the stack, and suppliers who can deliver that precision with verified, qualified products carry leverage that was not available to them in prior cycle.

The cost of validation rises because the interaction between components now matters in ways it did not at lower speeds.

The total margin is the system capability to absorb imperfections before losing data. At 1.6T, this margin is small. Each subsystem, oscillator, PCB, connector, laser, DSP, temperature, consumes a portion of it. When the sum of the portions exceeds the total available, the system fails.

Read more

The Memory Renaissance

For most of the last two decades, DRAM and NAND were treated like brutally cyclical commodities. When supply was tight, prices exploded. When fabs overbuilt, prices collapsed. Investors learned to watch inventory, capex discipline and spot prices, then move on.

That version of memory is not dead. But it is no longer the whole story.

Read more

The physics problem

Data center networks are migrating from 400G to 800G and then to the 1.6T standard, 1.6 terabits per second per port. At these speeds, data is transmitted using a modulation called PAM4, where each symbol carries two bits of information instead of one. To function, the receiver needs to sample the signal at a precise moment. This moment is defined by the oscillator.

In the 1.6T standard, each channel operates at 224 Gbps. The available interval for sampling a single bit of information is 8.9 picoseconds. A picosecond is one trillionth of a second.

Any time deviation within this interval is called phase jitter. And in the 1.6T world, this jitter needs to be measured in femtoseconds, which are a thousand times smaller than picoseconds. The best oscillators on the market today deliver jitter of 12 femtoseconds. For reference: light travels about 3.6 millimeters in 12 femtoseconds.

The transition from 400G to 1.6T is not linear. The jitter requirement drops fivefold while the data rate quadruples. This means that oscillators that worked perfectly for 400G are simply not suitable for 1.6T.

The clock for the physics problem

Quartz has dominated timing technology for decades for a fundamental physical reason: the piezoelectric effect. When an electric field is applied to a precisely cut quartz crystal, it vibrates at a mechanical frequency determined by its geometry and mass. This frequency is extraordinarily stable, far more so than any electronic circuit can reproduce on its own. The important fact is that the quartz used in high-precision applications is not natural. It is synthetic, grown in industrial autoclaves in a process called hydrothermal synthesis.

The process works like this. Inside a high-pressure steel container divided into two chambers, low-quality quartz chips are placed in the lower chamber along with an alkaline solution, heated to about 400°C. In the upper chamber, crystallographically oriented quartz crystal seeds are suspended at a slightly lower temperature, around 300°C.

This temperature gradient creates convection currents. The solution dissolves the quartz at the base, rises to the top where it cools and becomes supersaturated, and deposits pure silicon dioxide over the seeds. A growth cycle takes 40 to 90 days to produce high-purity quartz bars.

After harvesting, the bars are cut with diamond saws or lasers. The angle of the cut relative to the crystallographic axes determines the thermal stability characteristics of the final component. AT cutting is the most common for moderate temperature applications. SC cutting offers superior stability at extreme temperatures and is used in high-precision OCXOs.

The polished crystal receives gold, silver, or aluminum electrodes deposited by vacuum evaporation, is mounted on a ceramic base, and hermetically sealed under vacuum or nitrogen. The sealing is critical: any moisture molecule that alters the crystal’s mass causes frequency shift over time.

For frequencies of 625 MHz in fundamental mode, the crystal needs to be so thin that conventional mechanical cleaning techniques reach their physical limits. The solution is precision photolithography to sculpt the material, the same process used to manufacture semiconductor chips. Quartz crystals are the passive elements that define the base frequency: 156.25 MHz, 312.5 MHz, 625 MHz, they are the starting point of everything.

But the timing layer in a modern data center is not just the crystal. It’s a chain with multiple links, each adding precision or distributing the signal to where it needs to go.

Oscillators are the circuits that make the crystal oscillate stably and deliver a usable electrical signal. There are four relevant categories. XOs are simple oscillators, without temperature compensation. TCXOs have electronic thermal compensation that reduces frequency variation with temperature. OCXOs have an internal oven that keeps the crystal at a constant temperature, achieving stability on the order of 10⁻⁹, a level of precision used in critical networks and long-distance optical transmission. VCXOs have voltage control that allows fine-tuning of the frequency via an internal PLL. Companies worth noticing here: TXC Corporation, Rakon, SiTime

PLLs and clock generators are chips that multiply the reference frequency and generate multiple low-jitter outputs to feed different interfaces within the system. They are the amplifiers and distributors of the reference signal. I like to mention here: Renesas Electronics, Skyworks Solutions, Texas Instruments

Clock distribution buffers ensure that the signal reaches all ASICs and transceivers with the same phase, without degradation along the way. Same players as above.

Retimers act on data signals in high-speed links between GPUs and switches, extending the range and maintaining signal integrity through dynamic equalization. Here the big names: Broadcom, Marvell Technology

Network synchronization protocols such as SyncE and IEEE 1588 PTP ensure that different nodes in the data center network operate with the same time reference. Distributed AI networks may require holdover, the ability to maintain clock accuracy for hours without an external signal. Nokia appears here.

Innovation That Changes Everything: Fundamental Mode Without PLL

Historically, to obtain 625 MHz from a quartz oscillator, a 156 MHz crystal was used, and the frequency was multiplied internally with a PLL. The problem is that frequency multiplication amplifies phase noise. The resulting signal is faster, but dirtier.

The solution that is redefining the 1.6T market is the fundamental mode oscillator: the crystal resonates directly at the desired frequency, without multiplication. The signal is born clean.

River Eletec (6666.T), a specialized Japanese company, has developed the product that best represents this innovation: the KCRO-05, a 625 MHz differential oscillator that achieves jitter of only 12 femtoseconds using proprietary KoT-cut and OPAW technology. No internal PLL. The clock is generated directly at the operating frequency, with signal purity unmatched in the current market.

Nihon Dempa Kogyo (6779.T) another japanese managed to reach the same destination by a different route, using advanced photolithography to create 312.5 MHz and 625 MHz fundamental mode crystals with ±20 ppm stability, sufficient for long-distance transmissions where any frequency drift accumulates error over time.

MEMS: The Silicon Alternative

While quartz is grown in autoclaves for months, MEMS resonators are manufactured in standard semiconductor factories using silicon lithography, the implications are significant.

A MEMS resonator is up to 1,000 times smaller in mass than an equivalent quartz crystal. This mass difference has a direct consequence: immunity to vibrations and shocks. Server racks with high-speed fans and hundreds of mechanical components create a constant vibratory environment that can cause frequency jumps in quartz crystals. The MEMS resonator does not have this problem.

SiTime, a world leader in MEMS oscillators, developed Epi-Seal technology that encapsulates the resonator in an ultra-clean vacuum at silicon annealing temperatures, ensuring failure rates up to 50 times lower than those of conventional quartz.

But MEMS has a fundamental limitation that quartz does not yet have: floor jitter. SiTime’s best silicon Super-TCXOs achieve around 70 femtosecond, it is adequate for 800G, but still above the 20 femtoseconds required by pure 1.6T. For the most demanding optical transceivers, where every femtosecond defines the system’s margin of error, quartz remains the only option.

MEMS wins in reliability, speed of delivery, and integration while quartz wins in signal purity. That’s why the two coexist in the same data center: quartz in optical modules that need minimal jitter, MEMS in server and SmartNIC applications where robustness and flexibility matter more.

The companies that control this market

NDK — Nihon Dempa Kogyo (6779.T)

NDK is the most strategically positioned player in the industry. The company vertically integrates the entire process: it cultivates its own synthetic quartz in 118-ton autoclaves, processes the crystals internally, and manufactures the final oscillators. NDK’s EX-grade quartz has a purity of 7N (99.99999%), a level that analysts say no competitor will be able to replicate before 2028.

In November 2024, NDK launched the NP2016SBE and NP2520SBE oscillators specifically for the 800G and 1.6T markets, operating at 156.25 MHz and 312.5 MHz respectively. The products are already in production at major optical module manufacturers: Lumentum, Coherent, and AAOI. In May 2025, the company launched ultra-compact NH7050XA/NH9070XB series OCXOs for data centers and 5G base stations, operating up to 105°C.

NDK also supplies crystals for defense systems, including components for the F-35. The critical supply chain control that comes from this position is a real moat.

River Eletec (6666.T)

River Eletec is a smaller company, but it holds the most technically advanced product on the market for 1.6T. The KCRO-05 with 12 femtosecond jitter without PLL is currently the industry benchmark. The company directly serves OSFP-XD optical module manufacturers where femtosecond-level signal integrity is the competitive differentiator.

It’s the timing version of what River Eletec represents for 1.6T: small, specialized, technically ahead of larger companies in a critical niche.

Seiko Epson (6724.T)

Epson’s Microdevices division launched the OCXO OG7050CAN in October 2024 with 56% less power consumption (0.2 W versus 0.43 W in previous models) and 85% less physical volume. This was achieved through innovation in the rectangular SC cut of the crystal and the internal heating integrated circuit. Epson is positioning this product specifically for 5G backbones and data centers that require extreme stability with power constraints.

Daishinku — KDS (6962.T)

Daishinku launched the DS and DE series of high-performance differential oscillators. The 156.25 MHz DS series achieves 32 femtoseconds, a 65% reduction over previous versions. The DE series expands to 312.5 MHz with ±20 ppm stability for the 800G and 1.6T market. The company also maintains a partnership with SiTime for the distribution of MEMS oscillators under the KDS brand, positioning itself in both worlds simultaneously.

Kyocera (6971.T)

Kyocera announced in September 2025 the KC1210A series, an ultra-low voltage oscillator operating at 0.9 V — approximately 50% lower than traditional designs. The focus is on energy efficiency in high-density systems where every milliwatt saved in timing reduces the thermal budget for higher-consumption components. Samples began in 2025 with production scheduled for 2026.

TXC Corporation (3042.TW)

TXC is Taiwan’s leading volume supplier. It is not a specialist in ultra-low jitter OCXOs, but it dominates the TCXO and VCXO market for telecommunications and automotive with scale and certifications that few competitors can match. In data centers, TXC tends to appear in frequencies from 25 MHz to 156 MHz where jitter requirements are less extreme.

Siward Crystal Technology (8228.TW)

Siward launched 312.5 MHz differential oscillators with 30 femtosecond jitter in the OSC8A, OSC95, and OSCA5 series in 2026, targeting AI servers and 800G and 1.6T optical modules. It uses advanced photolithography to create ultra-small crystals in compact packages. It is the lowest-cost alternative in the high-performance range, positioned between leading Japanese players and the commodities market.

SiTime Corporation (SITM — Nasdaq)

SiTime is the disruptive player in the sector. In 2026, it acquired the timing division of Renesas Electronics for approximately US$1.5 billion, consolidating its position as a single supplier for many AI server designs. The company reported that 53% of its revenue in 2025 came from the communications, enterprise, and data center segments. In the fourth quarter of 2025, it announced a 50% increase in demand forecasts for 1.6T oscillators.

Its silicon Super-TCXOs achieve around 70 femtoseconds of jitter, sufficient for 800G and high-performance SmartNICs, but still below the 12 to 20 femtoseconds of the best pure 1.6T quartz products. SiTime’s strategy is clear: replace “legacy quartz” with programmable and flexible solutions that simplify hardware design and offer greater system reliability.

The Value Chain and Who Captures Margin

The timing chain has four links with very different economic dynamics.

Crystal manufacturers — NDK, Daishinku, Jilin — control the hydrothermal cultivation process. Raw crystal has low margins, but it is strategic. Whoever controls the high-purity quartz controls the entry point of the entire chain. NDK integrates vertically, from the crystal to the final oscillator, capturing margin at each step.

Oscillator manufacturers are where the margin increases. The crystal enters as an input and exits as a calibrated, tested, and certified product for a specific application. High-precision 1.6T oscillators cost many times more than simple oscillators, and the technical differentiation protects margins from price competition.

Module and ASIC OEMs — Lumentum, Coherent, Broadcom, NVIDIA — purchase oscillators as part of the bill of materials (BOM). The cost of the oscillator represents less than 3% of the total cost of an 800G or 1.6T optical module. Most of the margin is in the lasers, modulators, and ASICs. But winning approval from a module manufacturer is crucial: once qualified, the oscillator remains in the product for years.

Data center operators capture value at the solution level. They have little influence over the unit price of oscillators, but their demand forecasts affect average selling prices (ASPs) across the market. As SiTime noted, the increased forecast for hyperscalers to 1.6T was the main driver of the jump in average selling price in 2025 and 2026.

There is an important pattern here: each technological leap increases the average price of oscillators. The migration from 400G to 800G required higher frequencies. The migration from 800G to 1.6T requires 625 MHz and jitter below 20 femtoseconds. SiTime explicitly stated that “migrating to 1.6T requires higher clock frequencies, for which we obtain higher average selling prices.” The sector’s revenue growth comes not only from volume but also from the increase in the average price per unit.

Is it easy to change the supplier?

The answer is no. And that’s one of the reasons why qualified suppliers have staying power.

Many SiTime MEMS oscillators are designed to be pin-compatible with traditional quartz packages, allowing for direct physical replacement. But that doesn’t mean the swap is simple. Any change to the reference clock requires a complete revalidation of the system’s jitter budget, potentially affecting DSP and SerDes performance, which may require adjustments to firmware and signal equalization parameters.

For an optical module manufacturer that has spent months validating the system with a specific oscillator, the swap barrier is real and costly.

The lifecycle of quartz oscillators is also long, like decades, meaning that a win design today guarantees recurring revenue for a long time. The service and replacement business is an important part of manufacturers’ long-term economics.

The market and geopolitical risks

The global frequency control market, the full stack of crystals and oscillators, was estimated at around US$6 to 7 billion in 2026. The specific segment of oscillators for AI data centers and high-speed networks is a smaller fraction, but with accelerated growth and increasing ASP.

The main driver is not unit volume. Each data center switch uses few oscillators, not hundreds. The driver is ASP: the higher the required frequency, the more expensive the component. And the faster the data center needs to operate, the more expensive each oscillator needs to be to deliver the necessary jitter.

With 1.6T entering volume deployment in 2026 and 3.2T on the horizon, the industry average ASP is structurally up.

The production of ultra-high purity quartz is concentrated in a few suppliers, predominantly Japanese. NDK, Daishinku, Epson, and Kyocera control the critical processes. Any restriction in the supply of raw materials or in autoclave capabilities can cause delays that propagate throughout the optical module chain. At the peak of AI data center construction in 2024 and 2025, lead times for critical 800G components reached 50 weeks. The bottleneck was not in the lasers or network chips. It was in the oscillators.

This risk is one of SiTime’s main arguments for adopting MEMS: because silicon oscillators use the massive infrastructure of semiconductor foundries, the supply chain is more resilient and scalable than quartz production lines that rely on months of controlled growth in the lab.

The Future: Three Scenarios

Scenario 1 — Quartz Dominance (probability high)

Continuous innovations in miniaturization and energy efficiency reduce the historical disadvantages of quartz. Low-power OCXOs like Epson’s and ultra-small packages like NDK’s make the crystal competitive in dimensions that were previously MEMS territory. 1.6T and 3.2T modules continue to use high-purity quartz. The ASP of oscillators continues to rise as the required frequency increases. Crystal brands maintain high inventories, a sign that is already visible in recent NDK data.

Who captures: NDK, Daishinku, Epson, River Eletec. The Japanese oligopoly of ultra-pure quartz remains protected, especially if geopolitical restrictions limit Chinese suppliers’ access to the Western market.

Scenario 2 — Rise of MEMS (probability medium)

Technical advances reduce MEMS jitter to below 20 femtoseconds. SiTime, strengthened by the acquisition of the Renesas division, consolidates its position as a single supplier for AI server designs. On-chip timing solutions within optimized ASICs achieve comparable jitter for 800G and 1.6T. Quartz recedes to ultra-high frequency niches and defense applications.

Who captures: SiTime/MegaChips and system integrators. Bargaining power migrates from the crystal sector to the semiconductor sector. The Daishinku-SiTime partnership for the sale of MEMS under the KDS brand is an early indicator of this movement.

Scenario 3 — CPO Disruption (probability low)

Co-Packaged Optics matures faster than expected. Laser, modulator, and clock are integrated into a single photonic package. Timing migrates in-chip, generated by low-jitter PLLs manufactured on the same wafer as the optical components. Discrete quartz oscillators and MEMS become niche components.

Who captures: photonic silicon and optical chiplet manufacturers, Broadcom, NVIDIA, Ayar Labs. NVIDIA is already developing transceivers integrated into the same silicon package. In this world, discrete external oscillators are gradually eliminated, and the value migrates completely to whoever controls the optical chip.

Unit Economics: What Each Technological Leap Does to the Price

The global market for frequency control was estimated at $6 to $7 billion in 2026. The specific segment of AI data centers is a fraction of that, but with a price dynamic that has no equivalent in other components.

The logic is simple: each high-performance data center switch or ASIC uses between 2 and 4 specialized oscillators, one for the optical transceiver clock, another for the internal PLL. With tens of thousands of large switches in operation globally and unit costs ranging from $2 to $20 depending on the specification, the direct TAM of oscillators for AI and data centers is in the range of hundreds of millions of dollars annually, growing directly with the number of 800G and 1.6T optical ports being deployed by hyperscalers. The most important point is not the size of the TAM. It’s the ASP.

A simple 156 MHz XO costs a few dollars. A high-precision industrial OCXO for 312.5 MHz can cost tens of dollars. The migration from 400G to 800G to 1.6T doesn’t just increase the volume of oscillators. It increases the price of each unit because higher frequencies require crystals that are more difficult to manufacture, more precise cutting processes, and jitter tolerances that eliminate most suppliers. SiTime has explicitly stated: “the migration to 1.6T requires higher clock frequencies, for which we obtain larger ASPs.” Each technological generation is a mid-price upgrade for whoever can deliver.

Energy Impact - OCXO like Epson’s consumes about 200 mW. A complete 800G optical link consumes tens of watts. The oscillator is a small fraction of the total consumption. But with millions of ports in operation, every milliwatt saved in clock speed reduces the system’s thermal budget and allows for higher channel density without exceeding the heat dissipation limits of the racks. Kyocera’s 0.9V innovation and Epson’s miniaturized OCXO with 56% less power consumption are not just elegant engineering. They are a business advantage in data centers where every watt counts.

Conclusion

Timing is ceasing to be an invisible component and becoming a limiting factor in AI data centers.

Each speed increase in data center networks makes the oscillator more difficult to manufacture, more expensive, and more strategic. The 625 MHz quartz crystal in fundamental mode with 12 femtosecond jitter is not a component that just any company can produce, it now represents decades of expertise in crystal growth, cutting, processing, and hermetic encapsulation, all concentrated in a few Japanese suppliers with vertical integration and autoclave capabilities that cannot be replicated in years.

It is “crystal clear” that the physics of high-speed links demands increasingly precise clocks. The transistor will continue to shrink and the signal will continue to accelerate. And someone will need to keep up.