Google had the lead. DeepMind was years ahead in language modeling. They had the architecture. They had the data. They had the compute. But they didn’t launch. Not because they couldn’t. Because they wouldn’t. According to a senior insider, Google shelved its early LLMs to avoid cannibalizing its own ad business. A direct-answer model would have gutted Search. And Search is the engine that prints the money.
The logic was simple. A traditional search result sends users to websites. Those websites carry ads. Google gets paid. An LLM skips the click. It gives the answer directly. No link. No pageview. No ad. That’s not a feature. That’s a threat. So they waited. Then OpenAI moved. Then Microsoft backed it. Then the floodgates opened.
Now the shift is accelerating. LLMs are becoming the new interface for information. Instead of typing a query and clicking through ten blue links, users ask a question and get a synthesized response. It’s not just faster. It’s stickier. And it’s changing how traffic flows across the internet.
Marketers are already adapting. SEO is losing ground to AIEO. That’s AI Engine Optimization. The goal is no longer to rank on Google. It’s to be cited by the model. Agencies are rewriting content to be LLM-friendly. Structured data. Source credibility. Semantic density. The game has changed.
Some LLM providers are already monetizing the answers. Advertisers can pay to have their content favored in responses. Not directly. But through prompt tuning and source weighting. It’s early. But it’s happening. And it’s not sustainable on subscriptions alone. The compute costs are too high. The margins are too thin. Someone has to pay.
That’s where consolidation comes in. The companies that own social platforms have the edge. They control the data. Not just public web content. Private messages. Comments. Clickstreams. That’s the fuel for next-gen models. Everyone else is scraping the open web and burning cash. The clock is ticking.
Then came DeepSeek. The open-source model that showed how bloated the incumbents had become. It runs leaner. Trains faster. Costs less. And it exposed the inefficiencies in the current stack. What used to take 50 million dollars in compute now costs five. That’s not a tweak. That’s a reset.
The sector is under pressure. Ad budgets are shrinking. Subscriber growth is slowing. Retailers are cutting spend. SaaS firms are pulling back. The pot is not growing. It’s shrinking. Only the firms with other cash engines will survive the squeeze. Everyone else is running out of road.
Layoffs have already started. Developer teams are being trimmed. Infrastructure is being consolidated. The GPU farms built in 2022 are sitting half-idle. The demand projections were wrong. The revenue curve is flattening. And the reckoning is coming.
Google still has the scale. Still has the browser. Still has Android. But the moat is leaking. And the model is under stress. The question is not whether they can build a better LLM. It’s whether they can afford to let it run.
Sources
https://www.deeplearning.ai/the-batch/chatgpt-and-other-llm-could-disrupt-googles-business
https://x.com/DarioCpx/status/1930114296601813302
https://explodingtopics.com/blog/llm-search
https://europe.ark-funds.com/2024/07/google-and-the-innovators-dilemma-will-llms-reshape-search
https://dev.to/ashinno/best-code-llm-2025-is-here-deepseek-1e3m
https://www.deeplearning.ai/the-batch/deepseek-v3-redefines-llm-performance-and-cost-efficiency
https://bluetext.com/blog/social-media-advertising-whats-changing-in-2025