The Falling Cost of AI Favors Software Companies

By Kyle Levins, Analyst | August 20, 2025

Seven months after Chinese startup DeepSeek rattled markets by introducing a powerful open-weight large language model—freely available for anyone to run or modify, and built for a fraction of the typical cost—OpenAI has followed suit. In an apparent bow to the competitive pressure, the US-based company released two open-weight models on August 5, marking another pivotal moment in the economics of artificial intelligence (AI).

Just a year ago, integrating generative AI into software products was expensive, unreliable, and largely experimental. Today, it’s becoming both practical and profitable. As AI model costs continue to fall, software companies are weaving the technology into more of their products to improve customer productivity and create differentiating features that can command premium prices.

Three forces are driving down AI costs: a venture capital–fueled race for scale, the rise of open-weight models, and rapid advances in the underlying hardware.

Large language models are the brain of AI software products—they interpret language, generate content, and automate reasoning. OpenAI, Anthropic, and xAI, with the aid of massive investment by venture capital firms, are among companies racing to build the most capable and efficient models to attract customers and win market share. Together, the three have raised an unprecedented US$80 billion in only a few years. (For context, a total of US$200 billion was raised across the broader venture capital industry in the US last year.) That capital is being used to invest aggressively in workers—in some cases paying over US$200 million for a single engineer—and in techniques that push their models to be faster, cheaper, and more powerful for users.

A line graph titled 'Fundraising Frenzy' displays fundraising totals over time for OpenAI, Anthropic, and xAI, with OpenAI peaking near US$60 billion in March 2025.

Source: Company press releases.

Partly spurred by the innovations from China’s DeepSeek, there has been a Cambrian explosion in open-weight AI models in recent months. Unlike closed-weight models, users can customize open-weight models and run them on their own infrastructure. Although they are free for commercial use and redistribution, open-weight models increasingly rival closed-weight models in performance on popular benchmarks, such as the MMLU leaderboard, pressuring the ability of leading closed-model vendors to charge premium prices for their proprietary systems. OpenAI, the leading provider of closed-weight models that was once firmly committed to keeping its models proprietary, now appears to have shifted course.

A line graph titled 'Open-Weight Models Are Catching Up' compares the performance of open-weight (orange) and closed-weight (purple) models on the MMLU benchmark from October 2019 to July 2024.

Source: Epoch AI.

The key hardware that powers these models, graphics processing units (GPUs), is also rapidly getting cheaper on a cost-per-performance basis. Companies such as NVIDIA and AMD are releasing more powerful and specialized chips that reduce the cost of training and running AI models. NVIDIA’s latest Blackwell chip, for example, offers a thirtyfold improvement in speed over its H100 predecessor for AI use cases. These forces have combined to lower the cost of AI model tokens—the base unit of AI models’ input and output—by 75% in just one year. Features and products that were previously uneconomical are now viable. For example, in 2023, the leading workflow software vendor ServiceNow launched its popular ServiceNow Pro+ product that includes an AI-powered IT help-desk agent (and commands a 30% price increase relative to its Pro product).

A line graph shows the median price per 1 million tokens from OpenAI and Anthropic decreasing from US$30 in September 2023 to nearly US$0 by March 2025.

Source: Ramp.

Technology cost deflation is not a passing trend: it has long shaped software development. Most recently, it was true for cloud infrastructure. If AI models are the brains of modern software, cloud infrastructure is the bones. Leading cloud-infrastructure providers such as Amazon Web Services and Microsoft Azure offer the basic building blocks, including storage, networking, and computing power, to stand up applications. Thanks to steady advances in chip technology, as long described by Moore’s Law, the cost of cloud storage and computing power has fallen to near zero.

Over time, this decrease in cloud computing costs has been passed on to software companies, whose gross margins average 72%, higher than every industry except real estate investment trusts. It’s lowered the cost to create new software products that improve customer retention and deliver better value. For example, it allowed ServiceNow to enter new markets including field service management and for Adobe to launch mobile versions of its popular Photoshop software and new products such as Adobe Express. All this high-margin innovation has led to strong returns for software stocks. As of the end of 2024, software accounted for 10% of the total market capitalization of the S&P 500 Index, more than doubling from 4% in 2014.

AI looks to be following a similar path as cloud computing: While AI model builders compete aggressively and technology pushes forward, it is their customers, the software providers, that stand to benefit.

What did you think of this piece?

Is AI a Threat to Google Search?

By Uday Cheruvu, CFA, Analyst and Portfolio Manager | June 18, 2025

For the past two decades, the primary way of looking up information on the internet has been to Google it. But what about in the future? Will we ChatGPT it? Claude it? Will some other brand of artificial intelligence (AI) develop into the neologism of the coming era?

As people have begun to use AI chatbots in their personal and professional lives, it has introduced the possibility that someday Google search could become obsolete. For instance, Apple executive Eddy Cue said last month that he sees AI applications eventually replacing standard search engines, citing an unusual decline in search activity on Apple’s Safari web browser. Therefore, he said, the iPhone maker is working to add AI search options to Safari, where Google has long been the default search tool. Cue’s comments, which came during testimony in a federal antitrust case against Google’s parent Alphabet, caused Alphabet’s stock to plunge over 7% on May 7.

Source: Datos, SparkToro.

It’s common for commercial searches to draw on Google’s vast amounts of real-time data—e.g., what pizza shops are currently open, how busy they are, and customer reviews. This is a key reason AI chatbots haven’t yet made a dent in commercial search. The datasets used to train AI models aren’t constantly updated, and so they still need to be connected to a search engine to provide the most accurate and relevant live data. For now, AI chatbots are generally better suited to non-commercial searches, such as informational searches, which contain fewer advertising opportunities—for example, “what temperature to bake a pizza.” (However, Morgan Stanley’s survey did find that ChatGPT is gaining ground in travel research, particularly among teens and young adults.)

Over time, these distinctions may become less relevant, especially if people become more habituated to the interaction style and output of AI chatbots. However, it’s not as if these chatbots are on the verge of eating Google’s lunch and Google is standing by doing nothing. Both Google and its parent company are helping lead the way in AI-powered search with their own large language model, Gemini. For example, there are now Gemini-generated summaries at the top of Google search results as well as an AI search mode that can respond to natural language prompts and more complex questions. The company is also incorporating AI image-recognition capabilities into its products. In fact, Alphabet is the only true full-stack AI player because it has proprietary hardware and infrastructure (including chips and data-center technology), software (such as the Android mobile operating system and Google Cloud networking software), one of the most advanced large language models, as well as AI products for users across its ecosystem. Amazon, Meta, or Microsoft can’t offer such a comprehensive set of products.

Although there’s the risk that Google’s AI features will cannibalize its search ads by discouraging clicks to other web pages, the company’s future success depends primarily on one thing: maintaining the utility of a Google search. That means being willing to disrupt its own business to address changing user habits. While the search engine started out as purely an intermediary connecting user intent with other web sites, increasingly Google is not an intermediary but the destination itself. A user can often get all the information they need right on the Google search results page. AI is now supercharging that strategy by giving people even less of a reason to click away. For example, when searching “what temperature to bake a pizza,” Google’s AI overview helpfully suggested different temperature settings depending on the thickness of the crust and advised preheating the oven first and watching for “a golden-brown crust and bubbly cheese”—eliminating the need to scroll through results or read recipes for detailed instructions.

Where Alphabet faces a potentially more serious and immediate challenge is its pending antitrust litigation. The company is currently defending itself in two major federal lawsuits, one targeting its search engine and the other its digital ad business. In the worst-case scenario, Alphabet could be forced to share its search index—the underpinning of the Google search engine—to competitors at a nominal cost. If that were to happen, the collapse of Google’s search dominance could be sudden. In the advertising-related case, the judge could rule that Alphabet needs to share its user data for free with advertisers and publishers. However, both of those outcomes would be extreme and seem highly unlikely.

As for AI, the jury’s still out. But it has been more than two years since ChatGPT first arrived, and there aren’t any signs of Google’s imminent demise. Instead, what’s become evident is that Alphabet has a lot of compelling AI technology, and that Google search remains well ahead of anything else.

What did you think of this piece?

DeepSeek, Weak Outlooks Drag on IT Companies in Emerging Markets

By Pradipta Chakrabortty, Analyst and Portfolio Manager | April 22, 2025

High-quality stocks had a very poor first quarter in Emerging Markets (EMs), primarily due to Information Technology (IT) stocks, which also tend to trade at higher valuations. Stocks in the top third of quality underperformed the lowest third by close to 15%, the kind of sharp sell-off last seen four years ago when companies were still coming out of the depths of the COVID-19 pandemic.

Within the IT sector, semiconductor and other hardware stocks linked to the AI value chain sold off sharply after Chinese AI-startup DeepSeek released a large language model that rivaled the leading models at a fraction of the training cost, raising concerns that AI-related capex spending may drop.

Additionally, shares of IT-services companies were also weak. Most reported fourth-quarter earnings in line with expectations but provided disappointing guidance for 2025, quashing hopes of an acceleration in demand this year. Of course, additional trade-related uncertainty has skyrocketed since the end of the first quarter as global markets convulsed sharply in the wake of Trump’s tariffs in early April (with North America being a large market for many EM-based IT companies).

Still, despite the short-term uncertainty, the advancement of new technologies has increased the long-term opportunities for high-quality and fast-growing EM IT companies.

What did you think of this piece?

DeepSeek Rattles Markets But Not the Outlook for AI

January 31, 2025

A one-year-old artificial-intelligence (AI) startup born out of a Chinese hedge fund released a powerful AI model on January 20 that is challenging investors’ assumptions about the economics of building such systems.

R1, as the model is called, is an open-source, advanced-reasoning model—the kind that is designed to mimic the way humans think through problems. It was developed by DeepSeek, whose founder, Liang Wenfeng, reportedly accumulated 10,000 of NVIDIA’s graphics processing units (GPUs) while at his quantitative hedge fund, which relied on machine-learning investment strategies. The kicker: DeepSeek says it spent less than US$6 million to train the model that was used as a base for R1—a fraction of the billions of dollars that Western companies such as OpenAI have spent on their foundational models. This detail stunned the market and walloped the share prices of large tech companies and other parts of the burgeoning AI industry on January 27.

The knee-jerk reactions are a reminder that we are still in the early stages of a potential AI revolution, and that no matter the conviction some insiders and onlookers may seem to have, no one knows with certainty where the path will lead, let alone how many twists and turns the industry will encounter along the way. As more details become available, companies and investors will be able to better assess the broader implications of DeepSeek’s achievement, which may reveal that the initial market reaction was overdone in some cases. However, should DeepSeek’s claims that its methods lead to dramatic improvements in cost efficiency be substantiated, it may actually bode well for the adoption of AI tools over time.

Although the techniques employed by DeepSeek are not entirely new, the company’s execution is considered extraordinary for how it significantly reduced training and inference costs. The company reportedly not only used fewer and less-powerful chips but also trained its model in a matter of weeks. There is considerable debate regarding the accuracy of the roughly US$6 million training cost figure that has been reported as well as speculation that DeepSeek may have utilized rival models such as Meta Platforms’ Llama, Alibaba’s Qwen, and OpenAI’s GPT-4 in a process known as distillation to create a student model. Therefore, it may be that the true training cost of DeepSeek’s model is understated.

There is the risk that more efficient AI models may reduce the growth rate of demand for computing power and, in turn, the pace of investment in AI infrastructure. But while investors seem to have equated improved efficiency with a diminished need for AI infrastructure, the history of technology shows that better performance and lower costs typically lead to wider adoption and faster growth over time. AI could follow the same trajectory.

Among the likely beneficiaries of cheaper, more efficient AI are the software providers building applications on top of these models. As more users discover the advantages of AI, increased demand for user-facing software also helps the cloud-services companies that provide the computing power and data storage. However, the commoditization of large language models may negatively affect creators of these models, particularly those that have eschewed open-source technology.

Also, greater computing efficiency may lead to higher volumes of cloud data and thus the need for more data centers and electrical power over time. This would benefit infrastructure providers on the condition that higher demand more than offsets price declines.

The field of AI is still evolving, and further advances in the technology and the methods for developing it are likely. Innovation is always disruptive, but on the whole, DeepSeek’s breakthrough would seem to be a good sign for the industry.

What did you think of this piece?

Out of Our Minds

The Falling Cost of AI Favors Software Companies

Source: Company press releases.

Source: Epoch AI.

Source: Ramp.

Is AI a Threat to Google Search?

Source: Datos, SparkToro.

DeepSeek, Weak Outlooks Drag on IT Companies in Emerging Markets

DeepSeek Rattles Markets But Not the Outlook for AI

Disclosures