- The Angle
- Posts
- BYOT is the key to AI-native open source
BYOT is the key to AI-native open source
The mainstream view is that massive foundational model companies will eat the majority of the software industry. As mega-models get smarter, they’ll devour the application layer and leave traditional SaaS in the dust. But things might not play out that way. In fact, we might be entering a golden age of open source when we least expect it.
While frontier models are incredible, three counter-trends are quietly shifting power back to builders and enterprises:
The maturity of open-weight models: The open-source community is consistently punching at the weight class of proprietary giants. Models like Meta's Llama 3.1 and Qwen3 prove you don't need a closed API for frontier-level reasoning.
The Small Language Model (SLM) boom: The "bigger is always better" era is over. Models like Microsoft's Phi-4-mini and Google's Gemma 2 demonstrate that sub-10-billion parameter models—distilled and trained on highly curated data—can handle most production workloads at a fraction of the cost.
Better, cheaper tooling: Inference frameworks like vLLM, retrieval engines, and advanced quantization techniques have made running powerful AI locally, or within a private cloud, incredibly accessible.
This suggests today’s prolific thin wrappers may soon be displaced by a paradigm resembling open-source, reinvented for the AI era.
Thin wrappers are a conflict of interest. Right now, the application layer is flooded with software acting as simple token resellers. They may boast massive top-line revenue, but the bulk of it is a direct pass-through to foundational model providers. Enterprises tolerated this inefficiency in the initial AI rush. Today, buyers are getting savvy. They see the vendor lock-in, the fragile margins, and the risk of building on opaque APIs that can hike prices or deprecate models overnight. Why should a developer, for example, trust Cursor or Lovable (or Anthropic or OpenAI for that matter) to dictate which model to call and how many tokens to consume?
When intelligence gets cheap and ubiquitous, the real economic moats become obvious. What doesn't commoditize is proprietary data. What doesn't commoditize are messy, specialized human workflows where tacit industry knowledge lives. What doesn’t commoditize is taste, wisdom, and the decisions around what to build and why.
This brings us to the era of BYOT: Bring Your Own Tokens. Instead of routing enterprise data through centralized third-party wrappers, organizations and developers are starting to take control of the inference layer as yet another commodity input. Under BYOT, software platforms provide the orchestration, UI, and workflow, but the customer plugs in the engine. You can bring your own API keys or, better yet, point the software at a locally hosted SLM running safely within your secure perimeter. At Angular, we are already building internal software this way - and backing companies that are making this vision of the future a reality.
This lays the groundwork for a new software paradigm: AINOSS (AI-Native Open Source Software). Here, harnesses are open, inference is a fungible commodity, and vendor lock-in is explicitly engineered out when it comes to the model providers. Customers control the execution environment, protect their data, and choose exactly what to spend on inference - and with whom.
While coupling open harnesses with BYOT poses monetization challenges for AINOSS companies, smart product design is poised to solve them. Modern enterprises don't really want to vibe-code their own software from scratch and maintain it in perpetuity. There is ample opportunity for modern AINOSS vendors to build genuine product stickiness without resorting to artificial lock-in. Enterprises won't pay for the inference, but they will pay premium SaaS subscriptions for Role-Based Access Control (RBAC), archiving, SOC2 compliance, audit logs, and pre-built integrations to legacy databases. Thoughtfully managed workflows, intuitive UIs, and deep enterprise integrations will drive value in the AI era just as they did in the SaaS era—but now, they must compete on their own merits rather than acting as a tollbooth for another company’s intelligence layer. For many use cases, BYOT isn't just a cost-saving trick—it's the only logical way for an enterprise to scale AI without handing over the keys to the kingdom. The tooling is here, the models are capable, and a new generation of software builders is tackling this opportunity head-on.
Gil Dibner
FROM THE BLOG
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.
WORTH READING
ENTERPRISE/TECH NEWS
The never-ending last mile. Box CEO Aaron Levie argues that even when AI agents handle 80-90% of a task, the remaining slice still requires genuine domain expertise to evaluate, correct, and ship. More importantly, he thinks the last mile won't shrink over time: as AI raises the ceiling of what's possible, customer and market expectations rise with it, expanding the complexity of the role itself. Worth reading for founders thinking about where human expertise actually lives in an AI-native workflow, and what that means for hiring.
Legal AI wars. Microsoft just released their new Legal Agent, a shot across the bows of Harvey and Legora. That’s just a few weeks after Anthropic made a push into the world of law with Claude Cowork for Legal (webinar here). And just this past weekend, entrepreneur Will Chen launched Mike OSS, an open-source competitor to Harvey and Legora (which he described as “thin wrappers” with great GTM teams) that he vibe-coded in 2 weeks.
OpenAI's revenue miss. WSJ reported that OpenAI missed its 2025 ChatGPT revenue target and fell short of 1 billion weekly active users, with CFO Sarah Friar reportedly warning internally about the company's ability to meet spending commitments. The counterpoint, buried near the bottom of the piece, is that Codex is growing fast and GPT-5.5's developer reception has been strong. The debate matters for the whole sector: if even OpenAI is struggling to convert benchmark leadership into revenue, the real question is whether agentic coding tools are where the monetization story actually plays out this year.
China draws a line on AI exits. Beijing blocked Meta's $2 billion acquisition of Manus, the AI agent startup, and has reportedly begun warning Chinese tech firms to reject US investment without explicit government approval. The move is notable less for the deal itself than for what it signals: China is treating its frontier AI startups as strategic assets, not just commercial ones.
HOW TO STARTUP
When the agent goes rogue. Founder Jer Crane published a detailed post-mortem after a Cursor agent running Claude Opus deleted PocketOS's production database and all backups in a single Railway API call, then wrote out a confession explaining exactly which safety rules it had broken. This is a predictable consequence of moving so quickly. In this case, Railway's tokens carried blanket permissions with no scoping, volume backups lived in the same blast radius as the data they're supposed to protect, and destructive API calls required zero confirmation. Worth a read if you, too, are dangerously skipping permissions.
Agent burnout. This post makes a simple but underappreciated point: AI agents shift the bottleneck from typing to decision-making, and the human brain has a much lower daily budget for the latter. The worry is that ambitious 22-year-olds will treat agents as a forcing function to just work harder, not realizing they're burning cognitive fuel faster than before. Worth flagging for founders managing young teams: the productivity gains are real, but the burnout risk is too, and it's harder to see coming.
HOW TO VENTURE
Ineffable. David Silver, the former DeepMind researcher behind AlphaGo, raised $1.1 billion at a $5.1 billion valuation for Ineffable Intelligence, a months-old lab pursuing superintelligence via reinforcement learning. He's not alone: multiple former DeepMind, OpenAI, and Meta researchers have raised nine-figure seed rounds in the past year for labs that are barely out of stealth. For early-stage investors, the question continues to be…what does "seed" even mean anymore, if a founding team’s pedigree alone commands valuations that used to required $10s-100s of millions in revenue?
PORTFOLIO NEWS
FalkorDB is the #1 ranked open-source framework on GraphRAG-Bench, outperforming Microsoft and LightRAG by over 14 points.
Groundcover’s Co-Founder and CEO, Shahar Azulay, shares in interview with Informa TechTarget how observability is evolving from a post-production tool into the source of truth across the entire lifecycle, from code to production and remediation.
PORTFOLIO JOBS
Reply