How AI Moats Accumulate: Glean, Anthropic, and OpenRouter

Glean: the work behind a search box

Deedy Das first encountered Claude at Glean by tagging it in a Slack channel—an interaction he found odd for a new language model. The enterprise problem was already familiar: find useful information across a company’s systems while preserving who can see it.

Das moved from Glean into venture capital at Menlo Ventures. In this Latent Space conversation, he argues that Glean’s enterprise-search work from 2019 to 2022 created lasting value before the category became fashionable. AI accelerated an existing search business, making it easier to bring an already developed system to market.

His explanation of Glean’s moat—a lasting competitive advantage—centers on accumulated implementation and customer-specific work. Enterprise contracts can expand across knowledge workers, while deep integration makes the product difficult to replace. Das’s optimism also comes with a financial interest: he says he remains happy with his Glean stock and cites a $7 billion valuation at the time.

Alessio raises two pressures on that advantage: - Model labs entering search: Claude’s enterprise-search connectors bring a model provider into Glean’s market. - Source vendors restricting access: limits on application programming interfaces, or APIs, can constrain access to customer data. Salesforce’s restrictions on Slack are the example.

Das answers the access question with a hypothetical company buying 100 Slack seats. Glean, he says, shows Slack results only to licensed users with the relevant permissions. Permission-preserving search does not replace Slack licenses in that example; better discovery might instead encourage more use of the source product. Vendors may still restrict access to favor their own competing products.

Glean’s reported thousands of integrations reduce dependence on any single provider, though Slack remains critical for many customers and widespread restrictions would be more serious. Das also reports customer frustration: buyers regard the data as theirs and want to use other products to access it. He expects some resolution but does not know how the dispute will unfold.

Against model labs, his defense is partly economic. Deep enterprise search requires sales teams, engineers working directly on customer deployments, and extensive customization. Even a large individual contract may contribute little to a multibillion-dollar revenue business. Being able to build search differs from being willing to support every customer’s last mile. Das doubts that labs will consistently assign enough people to that work.

Ranking and adoption need more than a connector

An acquisition can leave an enterprise with two Slack workspaces and duplicate, overlapping systems. One participant describes struggling to use Devin across two Slacks; another recalls addressing multiple workspaces at Glean. Connecting company knowledge becomes harder when the company has two of everything.

Ranking has its own constraints. Consumer search can improve by aggregating clicks, hovering, and time spent on pages. Enterprise search cannot simply reuse that approach: - Sparse behavioral feedback: even a hypothetical 10,000 employees making two to five searches daily is described as insufficient for comparable feedback-driven ranking. - Freshness-sensitive queries: employees often need recently changed information. - Job-specific needs: common searches such as benefits and payroll cover only a small part of the value. - Domain-dependent evaluation: the vendor may understand neither a customer’s query nor its candidate results well enough to judge their ranking.

These conditions call for different ranking signals: relevance depends on the customer’s work, while the vendor may lack both abundant behavioral feedback and the expertise to assess results. Attaching a language model leaves that underlying retrieval problem to solve.

Adoption presents a separate challenge. Slack spreads through communication because colleagues need to join the conversation. Search is largely a single-user activity. The purchasing argument here therefore emphasizes employees finding a tool useful and continuing to use it, rather than a precisely measured productivity return.

Glean’s reported tactics placed search inside established habits: - Earn the new-tab page: where a customer already liked its browser start page, replace its internal search bar. - Meet users inside native search: a Chrome extension could replace Google Drive’s search interface with Glean results, automatically filtered to Drive. Internal evaluations led the team to believe its search outperformed native alternatives, though no methods or scores are given. The adoption mechanism was to let users encounter Glean while doing something they already intended to do.

Anthropic: product invention and enterprise staying power

Menlo’s Anthropic investment began before the company had revenue. The investment-side account describes growth beyond expectations: roughly zero to $100 million in one year, then $1 billion in another, with a $9 billion projection for the year of the conversation. A claim of the fastest software growth ever meets immediate pushback. The figures do not distinguish recognized revenue from annualized run rate, and the $9 billion remains a projection.

Claude Code brought a useful agent into the terminal. One speaker describes it as the first such product they personally saw delivering end-user value after the chat era. Its unusual interface prompts an organizational hypothesis: strong researchers and engineers given freedom can discover products that conventional planning might miss. Talent alone is an incomplete explanation, however, because OpenAI and DeepMind also have strong teams. The proposed link between employee autonomy and Anthropic’s product direction comes with explicitly limited visibility into its operations.

Advertising raises a related attribution problem. Participants prefer Anthropic’s later thinking-themed branding to its 2024 campaign, while leaving open whether better models or better marketing changed the reception. A billboard joking about a boss demanding AI branding resonates because workers face a more practical ambiguity: an instruction to adopt AI does not specify useful work to do with it.

The enterprise market figures have a narrower scope. Numbers attributed to Menlo put OpenAI’s share at 50% in 2023 and 25% in mid-2025, while Anthropic’s rose from 12% to 32%. These measure surveyed enterprise spending on large language model APIs, rather than token volume or the entire AI market; no sample size is supplied. One interpretation is that customers now have more credible alternatives, even as OpenAI’s business grows.

Similar model capabilities can coexist with durable customer relationships. Enterprises may stay once a model meets their needs, especially after buying long-term compute or dedicated instances. Hobbyist developers switch more readily. For the investment decision, revenue, margins, growth trajectory, and credible new markets consequently matter more than an isolated market-share milestone. The question is how future models and products sustain growth.

When the model provider also owns the application

The value of additional intelligence depends on the task. One participant tentatively argues that common consumer requests—dishwasher advice or rewriting an email—are already handled well enough that model improvements alone may do little for adoption. Coding still has substantial room for improvement, making better coding capability a more direct potential revenue driver. That does not remove price pressure: customers still weigh cost against performance, even when they prefer Claude.

Enterprise software strategies span a continuum: - High customization, high price: extensive customer-specific work, with less scalability. - Low customization, low price: broad access and faster scaling. The respondent places Anthropic toward the second end. Independent applications can benefit both their developers and Anthropic if enough of their usage remains on Claude. Claude Code’s margins are not disclosed.

Platform ownership introduces a competitive tension. A marketplace can observe demand and enter a seller’s category itself; Amazon’s branded batteries provide the analogy. The tentative moat thesis rests on an asymmetry: a model lab may enter an application market more easily than an application company can become a frontier lab. Competition among model providers still limits their value capture, and both layers can thrive. This is a strategic possibility, not a statement of Anthropic’s plans.

Claude Code makes the tradeoff concrete. One participant argues that supporting a strong independent application could sell more tokens than building a mediocre first-party product. But applications offering several models can direct usage elsewhere. The response rejects a universal ranking of interfaces—some users prefer Cursor or Devin—and identifies another benefit of ownership: application usage can supply data to improve the application.

Application defensibility also takes time to develop. The respondent sees relatively few applications with a substantial technical layer above the model, while acknowledging that many are only a couple of years old. A lab’s engineering and distribution may make entry possible; whether entry is strategically wise remains a separate decision.

Talent needs opportunity; ecosystems need useful support

The career discussion describes academic competition in India as a major route to social mobility, then identifies two ways credentials can constrain later progress: - Employers keep judging the past: early credentials can outweigh the quality of current work. - People internalize rejection: missing an early opportunity can become a belief that further effort is pointless. Another participant adds that choosing a workplace that rewards good work matters alongside credentials. Both selectivity and luck affect that choice.

The Anthology Fund supports startups building with and around Anthropic. Keeping the fund outside Anthropic avoided staffing an internal corporate venture operation. The investment rationale also challenged a common incentive in corporate funds: selecting companies primarily because they consume the parent’s products can be a poor way to invest.

Its reported scope spans three groups: - Strategically important companies: potentially larger businesses at later stages. - Strong companies using Claude heavily: attractive investments already building on the model. - Early founders with high potential: teams that may use Anthropic models. No particular model is required. Reported checks range from $100,000 to $20 million; a small initial investment can establish a relationship before a possible lead investment in the next round.

About 40 companies had reportedly received funding, including OpenRouter, Goodfire, and Prime Intellect. Events connect founders with Anthropic’s founders and executives. As Anthropic became better known, however, basic access to information became less distinctive, leaving the program to develop more useful support. The claimed advantage in reaching subsequent financing rounds is not quantified against a defined comparison group or period.

Research bets: understand models and distribute training

Research investing forces a difficult allocation decision: continue research toward a better product, or commercialize what already exists. The framework offered combines a team’s strongest technical competencies with a ten-year hypothesis: identify something likely to exist, then assess whether the team has a plausible path to building it.

Goodfire’s thesis is that consequential decisions will require understanding model internals. If models influence loans, insurance, and legal matters, output benchmarks and explanations generated by the models themselves may be insufficient. Mechanistic interpretability examines model weights to understand why a model produced an output, with potential applications beyond language models.

GPT-4o’s period of sycophancy—excessive agreement with users—provides the example. Personalized behavior may evade keyword checks and remain difficult to capture in targeted evaluations. The respondent claims interpretability methods can detect sycophancy and other problematic tendencies, including attempted deception and persuasion, though no validation conditions or measurements accompany those claims.

The bottleneck identified is access to weights, rather than model size. The response cites work on large open models, including DeepSeek models, and says Goodfire could work with Anthropic without access to Claude’s weights. Commercialization plans remain private; the respondent says some ideas were already working.

Prime Intellect’s distributed-training bet faces communication latency between geographically dispersed graphics processing units, or GPUs. The investment-side response accepts that risk. Its upside thesis combines distributed training, recruiting strong researchers from other labs, and a broader vision for which distributed compute is only the first step. Rapid changes in AI make the eventual direction difficult to forecast.

OpenRouter: maintaining model access becomes the product

OpenRouter provides access to many models through a maintained gateway. Connecting them looks easy until someone must keep every integration working. Persistent differences create a reason to buy the service. Its distribution thesis is product-led growth: developers can access and try the product without a sales call. The landing page earns praise for immediately showing information developers want—which models people use.

The business faces three distinct risks: - Lower prices can outweigh token growth: a service taking roughly 5% of spending faces pressure if prices fall faster than consumption rises. - Evaluation may not lead to retention: enterprises can try models through the gateway and then settle on one elsewhere; hobbyists may churn. - Frameworks can bundle a competing gateway: a Next.js developer using Vercel’s AI SDK may naturally adopt its gateway, while OpenRouter lacks a corresponding application framework.

The defense rests on focus, existing momentum, and details that larger platforms may not prioritize: - Data-retention controls: a reported routing option restricts requests to services that do not retain user data. - Provider-level visibility: the same model can have different context windows, output quality, latency, and throughput across serving providers. Usage across providers supplies information that may strengthen the gateway’s advantage. Choosing a model alone does not fully specify the service a developer receives.

Public usage rankings also attract users: open-model enthusiasts share charts celebrating increased usage. But popularity reflects pricing as well as quality. Grok Code Fast was described as leading an OpenRouter ranking at the time while benefiting from free access.

Dictation, diffusion, and the window for a new approach

Wispr turns dictation into usable text with less correction. Its Mac interaction is to press the function key and speak; it can handle spoken revisions that retract one phrase and replace it with another. Its internal zero edit rate measures how often dictated text needs no subsequent editing. A speaker estimates the rate above 80%, without specifying the sample, period, or testing conditions.

The investment argument is that interaction quality and accuracy can make speaking to a computer comfortable enough to retain users. The respondent describes Wispr as more than simply a Whisper model: the product’s value lies in making the input method reliable in everyday use.

Diffusion models offer a possible tradeoff among latency, cost, and quality. The proposed market is high-volume language or code generation that can accept lower quality for faster or cheaper inference. Whether diffusion reaches frontier quality remains open. The numerical efficiency estimates lack a defined baseline and workload; the investment case is that a different operating requirement may sustain a valuable product before it reaches the frontier.

Code supplies the structural motivation. Programmers revisit definitions and dependencies in multiple directions; matching opening and closing brackets is a small example. The speaker imagines diffusion’s denoising process repairing partial problems in several locations at once, rather than resolving everything before emitting a final answer. The architectural details are not developed further.

The counterargument invokes the hardware lottery: accumulated infrastructure investment and years of transformer research may keep an alternative from reaching the frontier, even if it is technically promising. More broadly, technical merit cannot guarantee market success. One participant cites MCP, an interoperability standard across labs and models, as a good-enough specification whose clarity and Anthropic backing helped it gain momentum.

Good execution still needs an open market window. One participant compares a startup to a fast runner in a collapsing tunnel: speed cannot guarantee an exit. MosaicML prompts a consequential distinction. One speaker questions customer readiness for fine-tuning, citing model quality and customers’ lack of suitable data and expertise; another points to its attractive acquisition. A weak standalone opportunity can coexist with substantial strategic value to a buyer.

Capital can change the market—and still require a payoff

AI rollups—buying existing businesses with a plan to automate their work—introduce a valuation dispute. In the hypothetical, someone buys a human-operated business with $1 million in annual recurring revenue for $2 million, automates part of it, and seeks a $100 million valuation by labeling the same revenue AI revenue. This is multiple arbitrage: obtaining a higher valuation per dollar of revenue through a change in category.

The operational defense is that an acquisition supplies customers and domain expertise, both difficult for engineers to obtain. Automation might then reduce costs and improve scalability. Funding can also create reflexivity: belief raises capital and equity value, helping recruit people who may eventually make the original belief true. These are plausible mechanisms, not reported operating results.

Large funding rounds may deter competitors before a product is ready. At infrastructure scale, participants argue that commitments to chips, machines, land, and power increase enterprise confidence that AI will persist. Amazon’s work for Anthropic and Stargate illustrate institutional commitment, while leaving a separate question: will the resulting capabilities justify the expenditure?

One participant describes an OpenAI roadmap from roughly 2 gigawatts of compute capacity in the recording year to 30 through announced projects over the rest of the decade, followed by a plan to finance another 125. These are ambitions and plans, rather than installed capacity. The debate separates two justifications for expansion: - Serve demand with existing capabilities: already-useful models might absorb more compute even without a dramatic next-generation improvement. - Create demand through better models: if current workloads cannot justify the expansion, research must produce capabilities that generate much greater demand.

Weak marginal model improvements could break the research-to-demand loop. A participant cites $5 billion of OpenAI research compute spending against $2 billion of inference spending in the recording year, then proposes research efficiency as an opening for a challenger. The accounting basis is unspecified, and the expense comparison does not itself measure research productivity or financial return.

A related competitive argument concerns willingness to spend. One participant argues that Google’s desire to protect search margins made it reluctant to incur additional compute cost on every query, creating an opening for OpenAI. An incumbent’s existing economics can constrain the response to a challenger.

Coding agents: who still understands the code?

The closing discussion returns to a concrete task. A participant recounts an unverified blog post about a suspected recruiting scam: a LinkedIn contact sent a repository to clone, run, and edit. The candidate reportedly asked Cursor to inspect it and discovered a byte array encoding a link associated with stealing private information. AI helped inspect the code in this account, while delegating execution without inspection could increase exposure to similar attacks.

The worry extends to skill development. One speaker compares repeatedly asking an agent to fix a problem to a slot machine: another prompt might produce the answer, making sustained reasoning harder to resume. That may displace the struggle through which engineers learn. A third perspective emphasizes delivery pressure: tickets must close, pull requests must merge, and applications must work.

One proposed design assigns different roles to assistance and delegation: - Fast assistance for difficult work: respond quickly, support the engineer’s reasoning, and stay easy to ignore when unhelpful. - Asynchronous delegation for familiar work: handle routine tasks the engineer already knows how to perform. The objection is behavioral: a powerful delegation option may be hard to resist, even when preserving focused thought is the goal.

The fast-agent proposal becomes concrete by dividing coding into finding and reading the right files, then writing the right files. A heads-up display for code comprehension would improve access to relevant information while leaving writing and actions to the human. Its proposed evaluation target is how much it helps the engineer read and understand; no evaluation results are reported.

Visible code and an explicit acceptance step, as discussed for Cursor, may encourage supervision without resolving dependence. The final distinction is experience. An engineer working in a familiar domain can challenge unnecessary files or poor implementation choices. A first-year student may treat the same output as an example of how software should be built. Meaningful human approval requires enough judgment to disagree.