By Hessie Jones
Large Language Models (LLMs), since their introduction three years ago, have upended how software is developed. Vibe coding is here, and anyone with a good idea now can build solutions from prompts and create a working prototype in days, not months.
Pre-LLMs, the moat was the algorithm. Companies would spend years building their custom models from curated datasets and hardening model efficacy over time as behavioral data increased. This has been the secret sauce—the key that defined product market fit and made them highly investable.
We’ve entered a stage of commoditized intelligence, general-purpose models courtesy of big frontier labs, OpenAI, Anthropic, Gemini and Deepseek. For companies integrating these more advanced AI applications, they will need to rethink how to leverage these LLMs to not only create a competitive edge but also to sustain their business.
In a recent AI Readiness webinar with Abigail Dubiniecki, Data Privacy Strategist and Consultant, she offered that the “shiny new object is no more” and that investors are less impressed by simple wrappers around large language models adding,
“As startups integrate large language models… many founders are realizing that technical feasibility is not the primary constraint. The deeper constraint is whether what they are building is legally permissible, socially acceptable, and commercially defensible… know what you are building, the risks it creates, and the responsibilities you cannot delegate. “
We reached out to our network of seasoned investors and advisors who have seen the evolution from narrow AI to generative AI models. They share their thoughts on what has changed this time around and how they’re evaluating new innovations in this nascent environment.
Shambhavi Mishra is an Associate Partner at Antler, a global venture capital organization investing in early-stage AI-native software companies across Saas, fintech, climate and infrastructure-adjacent technologies. In a recent post, she agreed that those who build generative AI wrappers will “struggle,” and adds,
“We are entering a reset phase, and the noisy layers will shake out, the real infrastructure will solidify, and GenAI will start to feel less experimental and more like basic operating infrastructure. Antler is focused on disciplined capital allocation and back companies that can build efficiently and sustainably.”
Mishra predicts that those with data moats in cybersecurity, compliance, defense and infrastructure will fare better.
William Ma is the Senior Director of Investments at MaRS Investment Accelerator Fund. He says this,
“AI has lowered the barrier to entry, so early traction isn’t a strong signal anymore. When everyone looks good at the start, we focus on teams that can turn an initial wedge into something durable and have the pace to adapt as the market evolves.”
When asked whether investors want more visibility to the current issues like copyright, or compliance impediments that have been the byproduct of these LLMs, Ma noted,
“It’s not a binary shift. It’s more about whether founders have a clear view of how their business will evolve, including the risks they’re taking on. For example, you might build an AI voice agent today, but competition will intensify, and switching costs will likely decline over time. The key question is how you build durability, whether through deeper workflow integrations, proprietary data, or other forms of compounding defensibility.”
Stephen Beney is the Principal at Smart & Biggar, Canada’s leading legal firm dedicated to intellectual property. Gurbir Sidhu is a Patent Lawyer, also with Smart & Biggar. As lawyers who evaluate AI startups, they agree they are shifting their approach as Beney adds,
“Traditionally, value in AI companies flowed from narrow, proprietary algorithms and specialized models. Today, with intelligence models increasingly commoditized, investors are asking a different question: What part of this business will still hold value when the underlying models can be swapped out?”
Gurbir adds that from an intellectual property (IP) perspective the analysis generally falls into two categories: IP Value and IP Risk:
IP Value
“As models become easier to replace, the strongest indicator of defensibility is often not the base model itself. Instead, investors look for whether the company has built a durable moat around the model.”
He emphasizes that a company may hold meaningful value even when using replaceable models if it possesses:
(a) Exclusive or hard-to-replicate training or retrieval data
(b) Defensible data labeling
(c) Reproducible training and use-case pipelines
(d) Strong safety, reliability, and evaluation benchmarking
(e) Demonstrated performance advantages under real-world constraints (latency, cost per task, failure rates) built over known models
For startups building in-house models, investors will inquire whether models will deliver superior economics, reliability or outcomes, with Beney adding,
“Are these advantages difficult for competitors to copy? Patents can help, especially when they map clearly to product value, but in many AI businesses, a layered IP strategy trade secrets, contracts, copyrights, and data rights tend to be more practically valuable than a thin patent portfolio.”
Gurbir introduces the second category:
IP Risk
“AI-first companies also face distinctive IP-related risks, and investors increasingly test how seriously teams manage them. The primary issue is data provenance and IP infringement.” He adds, a company developing its own models should be able to explain:
(a) What data it trained on
(b) What was licensed
(c) What came from customers
(d) What mechanisms exist for deletion, anonymization, opt-outs, access control, and restricted use
(e) How it addresses risks of IP infringement in the output from its models
Other key IP risks they site? Gurbir adds,
“Open-source license compliance, copyright exposure from training and retrieval workflows, risks of model memorization or extraction, and application‑layer vulnerabilities such as prompt injection and data leakage.
Regulators are also moving toward more explicit expectations around copyright, transparency, explainability, lack of bias, governance, and secure deployment.”
Beney suggests that from narrow AI to general-purpose models there is now broader shift in defensibility,
“In traditional narrow AI, defensibility often comes from a specialized model or proprietary algorithm built for a tightly scoped task. In the LLM era, base intelligence increasingly resembles a purchased input, usually accessed through an API rather than built in-house. While the shift reduces the burden of model training, it introduces dependence on third‑party providers and weakens any moat that relies solely on prompting or a thin integration layer.”
The strongest LLM-era startups operating on commoditized models therefore build defensibility elsewhere, such as, through system design, proprietary data advantages, secure and reliable deployment, workflow ownership, and disciplined IP risk management. When these components are in place, the commoditization of base models does not erode value. In fact, it can amplify the value.”
Zeeshan Ali, Partner at Wittington Ventures echoes this for other sectors such as climate and healthcare, where “…defensibility isn’t IP or data alone. It comes from translating models into real-world systems where constraints, regulation, and feedback loops compound advantage.”
Shambhavi Mishra, from Antler, advises that playbooks are being rewritten and adding that it has been difficult to keep up with the pace of this tech acceleration, but also asserts that fundamentals must be there. She states the importance on the ‘depth of workflow, not just intelligence’:
“The bar is no longer ‘we use AI.’ What matters is whether your product becomes deeply embedded in a customer’s workflow and is hard to replace. Distribution and integration are stronger moats than the model itself.”
Real usage and outcomes, not just early revenue, she argues, are key: “
In AI, early revenue can be misleading. We look for signals like repeat usage, retention, and whether the product is driving meaningful outcomes (time saved, revenue generated, decisions improved).”
Giselle Melo, Altitude Accelerator, Entrepreneur-in-Residence and Managing Director at MATR Ventures, doesn’t stray for the variables that investors use today, and contends revenue is a point of validation, explaining that while proprietary tech and IP are critical to “asset value,”
“We [as investors] place significant weight on tangible commercialization. Secure active paid pilots or early contracts to prove product market fit. Real-world revenue signals carry more weight than theoretical potential, especially when assessing applied or native AI companies.”
Mishra shares that companies must build towards a sustainable system—not a feature,
“Many AI products today risk being point solutions or are just solving for a feature. The companies that stand out are those that are evolving into a broader system or platform within their category, something that compounds over time. For example, an AI tool that starts with automating one step in a workflow (e.g., summarizing calls) becomes much more valuable when it evolves into a system that captures all customer interactions, generates insights, and informs decision-making across teams.”
Melo, from MATR Ventures, reminds the company’s TEAM is the biggest differentiator for investors,
“In an era where core algorithms are increasingly commoditized, we invest in founders first —more than ever. Be prepared to articulate not just what you are building, but why your team is uniquely qualified to build this company. Demonstrate your perseverance, domain expertise, and ability to execute—because in a crowded market, execution and the ability to identify where your growth is coming from are the true moats.”
And in an environment that currently demonstrates an increasing fragmentation because of political headwinds, Melo reasons that startups must think globally from day one, especially outside of North America,
“Show us that you understand opportunities to scale internationally and that you are open to leveraging our advisor network to mitigate risk and accelerate growth. A global mindset signals an awareness of the importance to navigate complex, evolving markets and globes shifting power structures.”
If you are a founder and looking for AI guidance to evolve your business, please check out our AI Resources Page for insights, videos on Building with Agentic Agents and AI-Powered coding and more…
Altitude Accelerator offers startup founders one path to scaling their business. We offer targeted programs based on the stage of your business. Whether you’re looking to bring your MVP to market, commercialize, or do your first capital raise, visit our program page to accelerate your growth. If you are currently raising, please start your application at Altitude Angels.
