When to ignore — and believe — the AI hype cycle

Picture this: It’s 2002. You’re lucky enough to get your hands on a first-of-its-kind smartphone that lets you message anyone in the world. Life changing, right? In the early 2000s, BlackBerry, Nokia and Ericsson were among the companies dominating the cellphone market. Fast forward to 2007, and the debut of the iPhone changed everything and eliminated the previous market leaders.

The iPhone revolution teaches us that the earliest innovators during a tech hype cycle don’t always emerge as the long-term winners. In fact, most often they do not. As the AI hype cycle continues to ebb and flow and early-stage generative AI startups sit at lofty valuations, this is a crucial consideration for all founders and VCs. 

What caused the AI hype?

The debut of OpenAI’s ChatGPT kicked off an avalanche of momentum in the gen AI space. Since then, nearly every major big tech player has released its own version, and 92% of Fortune 500 companies have adopted the tool. At the same time, a plethora of “wrapper” startups emerged with offerings that build off of ChatGPT’s model. 

One factor that clearly contributed to the buildup is the human tendency to overestimate change in the near versus long-term. We’ve already seen backpedaling in predictions around AI replacing jobs. For example, in 2020, the World Economic Forum predicted that AI would replace 85 million jobs worldwide by 2025. But their most recent report notes that AI is expected to be a net job creator.

While AI’s disruption to the workplace is undeniable, the hype bubble grows when we expedite timelines. Again, previous hype cycles showcase the value in refraining from making such claims. Another example of this is when key neural network research led to major breakthroughs in speech recognition and computer vision in the early 2010s.  

One article in Popular Science asserted in 2013: “We should probably just accept the fact that we’re that much closer to the sentient-robot takeover,” epitomizing the hyperbole that typically feeds technological hype cycles. This is not to undermine the significance of the breakthroughs brought about by deep learning in 2012, but rather to say we can take notes from the past to understand today’s AI frenzy. Here we are 14 years later, the robots haven’t taken over but the devices we use every day have become more frictionless and productive.

How to determine when an AI startup is worth the hype

Given how frothy the current AI market is, there are several considerations when choosing where to place your bets. As with any gold rush-like moment, it’s natural to look for the picks and shovels for others to build things and experiment — or in other words, create horizontal tools and infrastructure solutions.

At the same time, one has to be mindful that a key difference now versus in prior platform shifts is the pace of evolution. Established tech incumbents and startups are transforming their technology platforms simultaneously and big technology platform providers are also displaying an incredible amount of agility in adapting. This translates into a much more rapid evolution of the build with gen AI stacks compared to what we saw in the early days of the build with the cloud. 

If compute and data are the currency of innovation in gen AI, we have to ask ourselves where are startups sustainably positioned versus established tech incumbents who have structural advantages and more access to compute (while a lot of foundation model companies have also raised enormous sums of money to buy that access).

Higher up in the stack, the opportunity in applications seems quite vast — but given where we are in the hype cycle, the reliability of AI outputs, the regulatory landscape and advancements in cybersecurity posture are key gating factors that need to be addressed for commercial adoption at scale.

Lastly, foundation models have achieved the performance they have due to pre-training on internet scale datasets. What still lies ahead to realize the benefits of AI is the ability to assemble large, high-quality datasets to build models in more industry-specific domains. It is becoming increasingly clear that the biggest differentiator is the quality and quantity of data that models are trained on — and not the models themselves.

Keeping regulation on your radar

Given the excitement and broad potential for transformation from gen AI and large language models (LLMs), regulatory bodies around the world have taken notice. Whether it’s President Joe Biden’s recent Executive Order, or the EU AI Act, startups need to have a plan for regulatory what-ifs. 

This doesn’t mean they need to have all of the answers, but founders must have assessed potential regulatory hurdles and their implications. We’re in the midst of copyright battles and governments taking a stance on what data can and cannot be fed to AI models. More of these cases are bound to unfold.

Understanding cybersecurity considerations

Like regulation, AI innovation is outpacing cybersecurity. Businesses need to be aware when their company data is at risk of exposure from insecure, gen AI. We’ve already seen massive hacks due to security issues with third-party software providers, which have prompted businesses to reevaluate how they vet vendors. Startups must keep business’ cybersecurity needs and reservations in mind. 

Gen AI is opening up new attack vectors and surface areas in the enterprise. From adversarial attacks, prompt injections, data poisoning, to jailbreaking how models are aligned, much still needs to be addressed to make deployment at scale safe, reliable and robust. AI-infused cyber tools will certainly be part of defensive strategy, but protecting AI itself is an emerging sub-sector in cybersecurity. 

AI founders raise green flags when they demonstrate proactivity around regulatory and cybersecurity considerations.

Why data determines startup destiny

The biggest factor in whether a startup will be able to stand the test of time, through the noise of a hype cycle, is its data. Startups must be in control of their data destiny to derive sustainable value. A better question than “what is your gen AI strategy?” is “what is your data strategy?,” because a company’s model is only as good as the quality of its data. Access to high-quality data draws a line between success and failure. How an organization acquires, prepares and extracts value from data and has a path to building a data flywheel, is a critical success factor.

The vast majority of enterprise AI projects stall because of the inability to harness and prepare the appropriate datasets in enterprise. Another wrinkle is that a lot of industry use cases won’t have the luxury of internet scale datasets to start with. At least in some situations, this presents an opportunity for synthetically-generated data to force-multiply whatever data organizations can access. 

This is an area that has been exciting for several years and continues to hold promise for breakthroughs that can create a feedback loop of synthetic data enhancing AI models. We are starting to see notable examples of this at the intersection of autonomous vehicle development, gen AI and simulation tools. We could see similar approach with more verticalized foundation models.

Where is the AI hype cycle headed?

It’s clear that gen AI innovation will continue to come in waves and software and APIs will continue to mature in compressed cycles. Whether it’s Sora, Claude 3, or GPT-5, we will continue to see bursts in excitement as models demonstrate significant advances in capability. Similar to previous hype cycles, we must reckon with the reality that while nascent technology may be incredibly promising, it doesn’t give us the full picture — and we can’t jump to conclusions about what the gen AI wave means for every industry. 

I’d argue that the researchers, builders and doers are who we should be listening to, to get a sense of where the industry is headed — and not necessarily VCs, who are frankly better at picking companies versus long term trend predictions. 

Samir Kumar is co-founder and general partner at Touring Capital.