Insight Partners Managing Director Deven Parekh used a Sept. 13 TechCrunch interview to defend a diversified investment strategy at a moment when enormous pools of capital are concentrating around OpenAI and Anthropic. His argument was made for venture portfolios, but it points to a broader business-development discipline: companies adopting AI need options, not just a favored logo.
The concentration is measurable. PitchBook data reported by Yahoo Finance showed that AI startups raised more than $407 billion in the first half of 2026, with about half going to OpenAI and Anthropic. KPMG separately reported that AI captured 81% of global venture investment in the first quarter, when several multibillion-dollar rounds shaped the market.
Insight has exposure to both frontier-model companies. Anthropic listed the firm among the investors in its $65 billion Series H financing announced in May. Parekh told TechCrunch that the calculus changes by stage: an early investor with board access and sensitive information faces different conflicts than a later-stage holder making a more financial investment. He also said Insight avoids direct competitors at the earlier stages and uses information-sharing restrictions.
That nuance is important. Diversification is not the absence of conviction. It is a way to preserve strategic range when valuations, product capabilities and customer needs are moving faster than the evidence available to evaluate them.
AI concentration is now an operating risk
A company can become dependent on an AI provider long before procurement labels the relationship strategic. Teams build prompts and evaluations around one model, developers adopt its tools, customer workflows begin to rely on its response patterns, and sales teams make promises that assume its pricing and availability will hold.
Each decision may be rational in isolation. Together, they can create a concentrated operating position: one vendor controls a critical layer of cost, capability and customer experience. Switching then requires more than signing a new contract. It may involve rebuilding integrations, retesting outputs, retraining employees and revising the product promise.
Capital concentration increases the importance of that question. The largest providers have resources to advance quickly, but scale does not eliminate differences in model behavior, governance, commercial terms or product direction. It can also make smaller suppliers dependent on the same foundation platforms, creating hidden correlation across a supposedly diverse vendor stack.
A portfolio needs rules, not random experimentation
Analysis: The useful translation of Parekh’s portfolio logic is not “buy every model.” Enterprises should define where optionality creates value and where standardization creates efficiency. A disciplined AI portfolio starts with business outcomes, then assigns models and vendors according to the risk, economics and performance of each use case.
High-volume customer service may prioritize predictable cost and latency. Regulated analysis may demand stronger documentation and human review. Creative development may benefit from testing several models because stylistic range matters. Internal research may require a different data policy than a public-facing product. One provider may lead in each category without winning all of them.
The operating model should make those choices reversible. Teams can keep evaluation sets outside a vendor’s platform, document prompts and policies, map which data reaches each service, and build an abstraction layer only where the expected switching value justifies the complexity. Procurement can track concentration by workload and spend rather than counting vendor names.
Business-development agreements deserve the same scrutiny. Exclusivity can bring distribution, credits or technical support, but the consideration should be explicit. If a partner asks for preferred status, leaders should know what measurable advantage they receive and which future channels, models or customers the agreement could limit.
Conviction should be earned in stages
Parekh also described a practice of making smaller initial investments and increasing exposure when companies produce evidence. That logic can improve enterprise AI adoption. A pilot should be designed to answer a commercial question, not merely demonstrate that a model can perform a task.
Useful gates include verified quality against a baseline, total workflow cost, employee adoption, customer impact, failure handling and the effort required to change providers. A vendor earns a larger role when the results improve and the company can explain why. The process turns expansion into an evidence-based decision rather than a response to market momentum.
Leaders do not need to predict which frontier lab will dominate. They need a structure that can capture progress from the winners without making the business fragile if rankings, prices or partnerships change. In this phase of AI, optionality is not indecision. It is a form of operating leverage.