Was 2025 A Good Or Bad Year for AI? (Cal Newport and Ed Zitron Break it Down) | Cal Newport

Looking back at the tumultuous events of 2025, it is difficult to determine whether the year was defined more by breakthrough announcements or the quiet popping of an economic bubble. It was a year where the narrative shifted violently from "superintelligence is imminent" to "how do we pay for all of this?" From the shock of DeepSeek in January to the underwhelming release of GPT-5 in August, the AI industry faced its first true reality check. In a recent retrospective, Cal Newport and technology reporter Ed Zitron broke down the year month-by-month, dissecting the gap between the marketing mythology and the operational reality of companies like OpenAI, Anthropic, and Google.

The following analysis explores the pivotal moments of 2025, examining how the industry moved from promises of autonomous agents to a defensive scramble for revenue, and why the financial fundamentals of generative AI finally came under mainstream scrutiny.

Key Takeaways

• The Efficiency Crisis: The release of the Chinese model DeepSeek demonstrated that high-performance models could be trained for a fraction of the cost of American counterparts, challenging the "bigger is better" capital expenditure strategy.
• The "Agent" Flop: Despite early 2025 being branded as the "Year of Agents," autonomous coding and workflow agents failed to deliver reliable economic utility, leading companies to pivot back to core chatbots by year's end.
• The Inference Wall: The industry discovered that scaling costs do not vanish after training; the "test-time compute" required for reasoning models caused inference costs to skyrocket, damaging profitability.
• Technical Plateaus: The release of GPT-5 and Gemini 3 revealed that the era of exponential capability jumps had ended, replaced by marginal improvements achieved through expensive engineering workarounds.
• Media Skepticism: 2025 marked the turn of the tide for tech journalism, which shifted from breathless hype to rigorous investigation into the financial viability of the AI bubble.

The Early Year Shock: DeepSeek and the Myth of Moats

The year began with a destabilizing event that many in the Western tech sector tried to ignore. In January 2025, the release of DeepSeek, a Chinese AI model, sent shockwaves through the markets. While American companies were conditioning investors to expect training costs in the hundreds of millions, DeepSeek was reportedly trained for approximately $5.3 million. This efficiency challenged the prevailing narrative that only companies with access to massive, bespoke data centers could compete at the frontier.

Ed Zitron notes that the industry’s reaction was largely defensive, focusing on geopolitical tensions rather than the technical reality that capital moats were evaporating. While the story was quickly "memory-holed" by the media, it planted a seed of doubt: if high-performance intelligence is a commodity that can be built cheaply, the trillion-dollar valuations of US tech giants might be built on shaky ground.

The Failed Promise of "Agents"

Simultaneously, early 2025 saw a concerted marketing push declaring the "Year of Agents." Executives at OpenAI and Salesforce promised that AI would transition from passive chatbots to active digital employees capable of joining the workforce. This narrative was essential to justify continued capital expenditure as the novelty of standard chatbots began to wane.

However, the reality of "vibe coding"—where agents attempt to execute complex, multi-step programming tasks—proved to be economically limited. While agents could prototype simple dashboards, they struggled with reliability and security in enterprise environments. By the end of the year, major labs had quietly deprioritized independent agents in favor of improving their core chat products, tacitly admitting that the technology was not yet autonomous.

The Mid-Year Pivot: Hardware and Hype

As the year progressed, the narrative shifted from software capabilities to hardware necessities. Nvidia’s CEO Jensen Huang took the stage in March to announce a fundamental shift in how the industry viewed compute. He clarified that the future wasn't just about training larger models, but about "inference"—the ongoing cost of running them. This was a critical pivot: it signaled that AI companies would not just need massive capital upfront, but would face perpetual, scaling operational costs.

The Return of Doom Narratives

By April and May, as technical progress appeared to stall, the conversation briefly reverted to "AI Doom" scenarios. Reports like the "AI 2027" forecast predicted superhuman intelligence and existential risks. Critics argue this was a strategic distraction. By focusing on science fiction scenarios of future omnipotence, AI companies could divert attention from immediate, tangible issues like copyright theft, environmental damage, and the exploitation of overseas labor for data labeling.

"If they actually wanted to talk about scary bad things that are happening, talk about the Kenyans who are training these models for $2. How about you go and talk about the theft that's happening? How about we talk about the environmental issues?"

This "doom washing" served a dual purpose: it kept the hype cycle alive by promising god-like power in the near future, while allowing effective altruists and safety researchers to fundraise on the premise of saving humanity, rather than solving the practical unprofitability of the current models.

The Climax: GPT-5 and the Technical Ceiling

The pivotal moment of 2025 arrived in late summer with the release of OpenAI's long-awaited GPT-5. For over a year, the industry had anticipated a "Oppenheimer moment"—a release that would fundamentally alter the economy. Instead, the launch was met with a collective shrug. While the model was marginally better on benchmarks, it lacked the transformative qualitative leap seen between GPT-3 and GPT-4.

The "Router Model" Problem

Behind the scenes, the architecture of GPT-5 revealed significant engineering bottlenecks. Rather than a single, exponentially smarter model, GPT-5 relied heavily on a "router" system that directed queries to specialized sub-models. While this improved performance on specific tasks, it introduced severe inefficiencies.

Critically, this architecture broke the ability to effectively "cache" system prompts. In previous iterations, context could be stored to save compute. With the router model constantly switching between different model states, the system had to reload instructions repeatedly. This increased the "overhead" of every query, making GPT-5 significantly more expensive to run than its predecessors without delivering a proportional increase in value. This technical nuance was largely missed by the general press but was the smoking gun for infrastructure engineers: the models were getting more expensive, not cheaper.

The Financial Hangover: A Year of Losses

By the fourth quarter of 2025, the narrative had firmly shifted from technological awe to financial scrutiny. The release of lackluster updates like Google's Gemini 3 and Anthropic's Opus 4.5 in November barely registered with the public, indicating market saturation.

More damaging were the leaked financial figures. Reports indicated that OpenAI had spent upwards of $8.67 billion on inference alone through September, against roughly $4 billion in revenue. Anthropic faced similar burn rates, paying billions to AWS and Google Cloud. The math became undeniable: these businesses do not benefit from traditional software economies of scale. In a standard SaaS business, adding a user costs almost nothing. In generative AI, every query requires firing up massive clusters of GPUs, meaning costs scale linearly—or even exponentially—with revenue.

Desperation Moves

The year ended with signs of strategic flailing. The launch of the Sora app—essentially a TikTok clone powered by expensive video generation—was viewed by analysts not as innovation, but as a desperate attempt to find a consumer use case for idle compute. Similarly, high-profile partnerships, such as Disney’s investment in OpenAI, appeared to be more about stock positioning and "innovation theater" than practical integration of tools into professional workflows.

Conclusion: The Bursting of the Bubble

So, was 2025 a good or bad year for AI? The evidence suggests it was a terrible year for the *business* of AI, but perhaps a necessary year for the *truth* about AI.

The "magical thinking" that dominated 2023 and 2024—the belief that scaling laws would indefinitely produce smarter models and that costs would naturally plummet—collided with physical and economic reality. The industry hit a wall in data availability, power consumption, and algorithmic efficiency. While the technology remains useful for specific tasks like coding assistance, 2025 proved that it is not a magic wand for the global economy.

As Ed Zitron concluded, 2025 was the year the "milk curdled." The glitzy promise of AGI has been replaced by the stark reality of massive burn rates, undelivered features, and a media class that has finally learned to do the math.