Arming the resilient enterprise in the age of AI

Recently, some of the key players in the AI world began questioning the dominant narrative. On June 30, 2026, Palantir posted a nine-point manifesto outlining private AI ownership, targeting the financial and operational risks of "tokenmaxxing." The next morning, Palantir CEO Alex Karp delivered a sharp broadside on CNBC’s Squawk Box, warning that enterprises are paying for "tokens that create no value" while transferring their intellectual property to closed-model providers. A few days later, on July 5, 2026, Mistral AI’s CEO Arthur Mensch voiced a parallel warning regarding "AI lock-in," detailing how heavy reliance on closed frontier models exposes enterprises to sudden pricing changes, service withdrawals, and structural vulnerability. Something is afoot: people are beginning to say the quiet part out loud.

At Angular Ventures, we have long believed that the initial phase of uncritical third-party AI adoption would inevitably shift toward first-party AI infrastructure, local/proprietary context, and strict data control. Given the choice between building with powerful new AI tools or depending on mercurial third-party foundational model companies, we think enterprises will choose the former. Consequently, we have focused our infrastructure investments on AI-native companies that are building the technical plumbing required to leverage models securely within the enterprise perimeter. We have rejected the premise that foundational models will simply eat the world and—instead—have bet on the tooling that enterprises (and AI-native software vendors) need to thrive in this new era.

We are thrilled that the market is beginning to catch up to this reality, and we wanted to share an overview of how we have positioned our portfolio to benefit from these trends:

1. Reco: Securing the AI and agent ecosystem. The rapid adoption of AI and autonomous agents creates a broad security footprint within the modern enterprise. We backed Reco (reco.ai) because organizations cannot safely deploy technology they cannot see or govern. Reco provides continuous discovery and visibility across SaaS environments where AI features and agents operate. By mapping identity, permissions, and data flows, Reco prevents autonomous agents from exposing sensitive corporate data or executing unauthorized workflows. This foundational security layer acts as a critical safety moat, allowing enterprises to adopt advanced AI tools without sacrificing control over their operational perimeter or exposing their internal intellectual property to external risks.

2. Groundcover: Private, AI-native observability. We backed Groundcover (groundcover.com) because its architectural approach solves a fundamental data bottleneck in modern enterprise monitoring. By using a novel kernel-level eBPF approach with a Bring Your Own Cloud (BYOC) architecture, groundcover enables cost-effective, default inspection of infrastructure telemetry without manual code changes. This results in a complete data corpus stored safely within the customer's secure cloud perimeter. Crucially, this high-fidelity dataset is optimal for AI-powered root-cause analysis and anomaly detection, transforming observability from passive dashboards into an automated engine. Pre-AI, no enterprise could afford to monitor everything. Post-AI, no enterprise can afford not to.

3. FalkorDB: Establishing ground truth and guardrails with ultra-fast GraphRAG. Deploying enterprise-grade AI requires more than just a model; it requires building rigorous guardrails, evaluation frameworks, and an authoritative source of truth. We backed FalkorDB (falkordb.com) because it provides the data foundation necessary for this reliability. FalkorDB is an open-source graph database designed to automate the creation of Knowledge Graphs for Retrieval-Augmented Generation (GraphRAG). By delivering performance up to $500\times$ faster than traditional graph databases, FalkorDB allows enterprises to establish a dynamic ground truth that anchors AI reasoning. This performance advantage ensures that AI systems can prevent factual errors and maintain compliance with strict operational guardrails in real-time, industrial-scale environments.

4. DataFlint: Navigating data infrastructure complexity and cost. Data infrastructure is costly and complex—challenges that a generic foundational model cannot handle in isolation. We backed DataFlint (dataflint.io) because optimizing these complex data environments requires real-time access to live enterprise queries and actual data models. No third-party model can address this problem - not now and not ever. The context is too large and changes too fast. DataFlint agentically tracks and optimizes enterprise data pipelines in real time. By putting performance intelligence directly into the hands of the enterprise, DataFlint ensures that data infrastructure remains performant and cost-effective without relying on generic, external cloud models to manage operational complexity.

5. Baseshift: Ensuring the stability of production databases in the agentic era. In the era of agentic AI, engineering velocity scales infinitely, but database logistics have become a key bottleneck. We backed Baseshift (baseshift.com) to eliminate that constraint. Baseshift provides scalable database logistics designed to secure production environments and accelerate developer workflows. Their "Baseshift Guard" technology protects production databases by intercepting rogue queries from autonomous agents before they reach the live environment. Simultaneously, Baseshift’s proprietary storage architecture allows teams to spin up lightweight, read-writable "Super Replicas" in seconds. This enables engineering teams to cost-effectively scale their database infrastructure to match the rapid, experimental output of agentic coding systems.

6. SpecificAI: Driving task-specific efficiency. Running large, general-purpose models for simple corporate workflows introduces increased latency, high costs, and unnecessary security risks. We invested in SpecificAI (specific.ai) because it solves this by making specialized first-party small language models (SLMs) available at true industrial scale. Its platform helps enterprises condense the complex behaviors of expensive, generalized agentic models into smaller, task-specific models that are optimized for performance and reliability. This approach reduces operational overhead significantly while enabling the deployment of secure, highly-optimized AI infrastructure within private cloud or edge environments. By focusing on specific utility rather than general intelligence, SpecificAI ensures that enterprises can deploy AI agents that are faster, safer, and far more cost-effective for high-volume production tasks.

7. Hornet: Re-architecting retrieval as a core AI stack layer. We invested in Trondheim-based Hornet (hornet.dev) because enterprise AI performance is ultimately bottlenecked by the quality of the information it can find. Hornet’s research highlights this gap: even high-performing models see accuracy drop from 93% to 14% when relying on standard, unoptimized search tools to find corporate data. This steep decline demonstrates that finding the right context is the primary challenge in enterprise AI maturity. Hornet addresses this by building a dedicated retrieval engine specifically designed to help corporate AI systems reliably search, organize, and access private business data at scale. By re-architecting retrieval, Hornet ensures that the "intelligence" of the model is always matched by the accuracy of the underlying context.

8. Our latest stealth investment: Enabling true "bring-your-own-token" agentic development. We recently invested in a stealth-stage company addressing the software developer environment. As enterprises build out internal AI agents, they risk becoming dependent on the proprietary SDKs and development frameworks of specific model providers. This team is building an open-source harness that enables true "bring-your-own-token" development. By decoupling the agentic build environment from the underlying model provider, this platform allows enterprises to switch models seamlessly, use their own private endpoints, and maintain control over their runtime environments without vendor lock-in. In their vision of the future, the models are commoditized and even the most sophisticated harnesses are entirely open.

Towards enterprise-grade AI maturity

The landscape Karp and Mensch describe is not a retreat from AI, but rather a transition towards enterprise-grade maturity. Enterprises are no longer content to let external foundation model providers dictate terms, collect their proprietary contextual data, or introduce structural lock-in. This is no longer about how fast a cool demo can get off the ground or who can spend the most on tokens. Instead, organizations are taking charge of their own technical destiny—and building rationally towards an AI-enabled future.

The seven companies highlighted above represent the first-party plumbing of this self-determined future. By decoupling core operational workflows from individual model APIs—via secure observation, robust private grounding databases, localized optimization engines, and open-source development harnesses—these technologies allow the enterprise to remain the sole beneficiary of its own data and intelligence. For early-stage founders building the deep infrastructure that powers this transition, we believe the opportunity has never been clearer.

Gil Dibner

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

Reverse Information Paradox - Satya Nadella this week wrote “In the age of intelligence, how should firms protect their core IP?” In an essay posted on X (and elsewhere), he argues AI inverts the classic economics of information: where sellers once risked giving away knowledge just by revealing it (Arrow's paradox), buyers of AI now risk giving away proprietary knowledge just by using it well — every prompt, correction, and interaction teaches the model provider more about you than you learn about them. His proposed fix is a hard "trust boundary" around enterprise data, traces, evals, and fine-tuned weights that nothing crosses without consent, built on five pillars (control, capability, choice, cost, compound) so firms own their own continuous learning loop rather than transferring that value to model vendors.

Fable v Sol - Forbes covers the latest in the AI model wars as Anthropic extended free access to Claude Fable 5 through July 19 for subscribers, a move that lands just three days after OpenAI launched GPT-5.6/Sol, which OpenAI claims beats Fable on coding and notes Fable "falls back" to an older model for most biology/chemistry queries. Fable is a safety-restricted variant of Anthropic's flagship Mythos model (currently limited to ~150 organizations, built for defensive cybersecurity use); Fable itself was briefly pulled by the Commerce Department last month after Amazon researchers found a jailbreak letting it walk through exploiting a software vulnerability, a gap Anthropic says is now blocked in over 99% of cases. The piece frames this against an increasingly public Anthropic–OpenAI rivalry (Amodei's comments on distrust of Altman, Altman's earlier "dishonest"/"authoritarian" jab over Anthropic's Super Bowl ad) and Anthropic's separate tension with the Trump administration, which designated it a supply-chain risk after it refused the Pentagon unrestricted model access. 

HOW TO STARTUP

Bad Vibe(s) - Wix CEO Avishai Abrahami sat down with Harry Stebbings on 20VC this week, and the core thesis is that Wall Street has mispriced Wix by overestimating the threat "vibe coding" poses to its core SMB platform. The stock is down roughly 71-83% from its peak and now trades around $50/share, with a ~$2.9B market cap, with a forward P/E of ~10.8, below the industry average of ~13. Analysts peg fair value closer to $78, so there's a real gap between the multiple and the narrative.

Abrahami's pushback: "You're not going to vibe code Shopify, no matter how good you are, the business stack is too hard," and separately, "JPMorgan will trust Salesforce with their customer data. They won't trust anybody else" — his argument being that trust and complexity in enterprise software categories aren't things AI code-gen erodes away, even as it gets better. More broadly, he says "we all give too much credit for AI and what it can do."

Wix isn't just defending against vibe coding, it bought into it. Base44, a one-person startup Wix acquired for $80M, and is now doing $150M+ ARR, giving Wix a direct stake in the trend regardless of which way it breaks.

HOW TO VENTURE

Opinionated Venture - Off the back of his move from Lightspeed to Union Square Ventures, where he joined as GP Mike Mignano shared his thoughts on 20VC about the current state of play in AI and Venture. His formula for seed investing: have a great network, ship your ideas, and be willing to take bets on people. On ‘ship your ideas’ if you’ve read this far on The Angle, you’ll know that we obviously agree. Mignano states: “To be great at seed investing, you have to put your ideas into the world. Sharing your thinking signals to founders what you believe in and what kinds of companies you're looking to back. Content isn't just marketing, it's a way for the best founders to find you before everyone else does."

DualBird
Senior Software Engineer (Haifa)

Firebolt
Senior Software Engineer (Munich)

Jurnee
Software Engineer Intern (Paris)