From underwear to unhinged: Inside the AI engine driving online advertising’s next boom
Jason Spilkin December 2025
During the National Basketball Association (“NBA”) finals in June this year, an upstart prediction exchange, Kalshi Inc, ran a national television advert which went viral. The “unhinged” ad included a farmer floating in a pool of eggs, an alien chugging beer, and an old age pensioner draped in an American flag screaming “Indiana gonna win it, baby” with the finale, “the world’s gone mad, trade it”. These surreal scenes were only possible due to Artificial Intelligence (AI) and Google’s Veo 3 text-to-video generator.
What made Kalshi’s advert dramatically different, is that it cost only $2,000 to develop, compared to six to seven figure quotes from traditional advertising agencies. It also took just three days from idea generation to finished product. Indeed, Kalshi’s advert creator was one guy, working from home, mostly in his underwear. Self-described “AI filmmaker”, PJ Accentura, used Google’s advertising platform and AI tools exclusively. First, he wrote a rough script and used Gemini’s chatbot to generate appropriate prompts. Then, he pasted the prompts into the Veo 3 text-to-video generator to get a rough cut. Finishing touches were made using Google’s editing software.
It all started back in 2012, when computer scientists at Google’s X lab built an artificial brain - a “neural network” with sixteen thousand processors and over one billion connections and let it browse ten million randomly selected YouTube videos. After three days, when presented with a list of twenty thousand different items, Google’s “brain” was able to recognise pictures of cats by using a deep learning algorithm. "Contrary to what appears to be a widely-held intuition, our experimental results reveal that it is possible to train a face detector without having to label images as containing a face or not”.
The cat experiment proved that, if one built a big enough brain and dumped sufficient data therein, it could learn unsupervised without any human intervention. “The idea is that instead of having teams of researchers trying to find out how to find edges, you instead throw a ton of data at the algorithm and you let the data speak and have the software automatically learn from the data”.
Google was early to realise that Graphics Processing Units (“GPUs”), which power gaming consols such as Sony’s PlayStation, were (paradoxically) more efficient at processing massive computational requirements of neural network applications, compared to traditional CPUs. In 2014, a decade before the AI frenzy took flight, Google placed a gigantic GPU order with Nvidia.
Not needing all the bells and whistles of the off-the-shelf GPU, Google started designing their own custom Tensor Processing Units (TPUs) which today are the only other scaled deployment of AI chips (besides Nvidia). TPUs are not as good at graphics and have poorer precision, but they are far faster and more energy efficient, giving Google’s Cloud Platform (“GCP”) a competitive advantage as a low-cost AI provider, which means that it can navigate around the “Nvidia tax”. Metaphorically, Nvidia GPUs are a Ferrari, whereas Google’s TPUs are a tuned-up Toyota.
With their vast vault of videos, photos and messages, online platforms such as Google and Meta are especially well positioned to offer AI tools. Indeed, there is ambiguity as to whether OpenAI improperly poached YouTube’s data to train Sora (text-to-video). In an interview with the Wall Street Journal, OpenAI’s Chief Technology Officer was evasive, saying she was “not sure” whether YouTube, Facebook or Instagram videos were used, eventually pivoting to say that “only publicly available, or licenced data” was being used. In an interview with Bloomberg, YouTube’s CEO confirmed that training AI models on YouTube data would be a breach of its terms and conditions.
AI also allows online platforms to more precisely target customers by tracking their trove of user data and making more appropriate advert recommendations. In search, Google noted this quarter that AI Overviews and AI Mode are improving the customer experience, driving greater frequency (i.e. number of searches) and complexity. Though the click through rate (per query) is likely lower, the offset is higher quality clicks (conversion rates), which monetise at a similar rate to traditional search. Last year, conventional wisdom was convinced that Google’s search revenues would be disintermediated by chatbots; however, a year on, the relationship seems more symbiotic than parasitic. Meta noted that AI is improving recommendation systems, driving more relevant content and boosting time spent on their platforms.
Both Google and Meta boosted AI investment guidance last quarter. Google’s capex boost was well received because the prospective returns are tangible today. GCP offers capacity to third parties and increased investments are being driven by insatiable appetite for AI capacity to run ravenous content creation tools such as VEO and Nano Banana (text-to-photo) in addition to internal uses (advertising targeting). The company noted, “We continue to invest in AI part features that are helping creators supercharge creation and build their businesses. With Veo 3 integration and Speech to Song, creators go from idea to iteration quicker, and new channel insights help them better understand performance”.
In contrast, Meta’s AI capex increase went down like a lead balloon. Unlike Google, Meta does not offer a cloud infrastructure platform for third parties. Rather, they are aggressively front-loading capacity buildout, seeking to vertically integrate their own AI capability. CEO, Mark Zuckerberg’s, zest for self-reliance probably has its origins in Apple’s privacy initiatives, which included App tracking transparency in 2021. Most Apple users opted for “ask app not to track”, thwarting Meta’s advertising tracking capabilities, resulting in a $10bn revenue hit in 2022. However, Meta only embarked along the road to “superintelligence” a decade after Google. Moreover, it lacks a third-party platform to realise economies of scale. Hence, we believe they would be better off licensing AI technology but leveraging their own user data (like Apple).
Indeed, Meta might end up there eventually. Shareholders can be forgiven for having flashbacks to Zuckerberg’s ill-fated metaverse foray, which they are reminded of every time they hear the name of the company. Thankfully, that calamity was corrected 12 months later when Zuckerberg cut capex to the bone, embracing a year of efficiency which ended in a triumphant 2023-24. Ultimately, he is no fool, has a lot of skin in the game and Meta’s capex binge is entirely discretionary. Historically, he has shown propensity to pivot where returns prove elusive.
In the 19th century, economist William Stanley Jevons observed that after the invention of the fuel-efficient steam engine, coal consumption unexpectedly increased. “Jevon’s paradox” occurs when technological advances make a resource much cheaper but spending nonetheless increases due to “price elastic” demand. In other words, modest changes in price lead to disproportionately bigger boosts to demand. Neither YouTube nor Facebook could have existed in their current form 30 years ago with dial-up internet which was slow and expensive. Both companies have benefited massively from cheaper, high-speed internet and ubiquitous smartphones which allow videos to be played in the palm of one’s hand.
History suggests that AI is another technological tailwind driving demand acceleration in online advertising. It is making content creation cheaper and more capable. That is why today one can watch viral videos, such as a recent one of US President Trump taunting “No Kings” protesters from an F-16 with the Top Gun tune “danger zone” belting out in the background. It is also improving customer targeting, such that the X (formerly Twitter ) algorithm knew who would find that clip comical. Both factors are boosting engagement and monetisation, accelerating growth in advertising revenue, boding well for both Google and Meta over the medium term. Jevon would have not been surprised.
This communication does not constitute an investment advertisement, investment advice or an offer to transact business. The information and opinions expressed in this communication have been compiled from sources believed to be reliable. None of Credo, its directors, officers or employees accepts liability for any loss arising from the use hereof or reliance hereon or for any act or omission by any such person or makes any representations as to its accuracy and completeness. Any opinions, forecasts or estimates herein constitute a judgement as at the date of this communication. Credo Capital Limited is a company registered in England and Wales, Company No: 03681529, whose registered office is 8-12 York Gate, 100 Marylebone Road, London, NW1 5DX. Authorised and regulated by the Financial Conduct Authority (FRN:192204). © 2025. Credo Capital Limited. All rights reserved.
