The evolving economics of AI, Part 2

The rapid emergence of smaller, cheaper and increasingly capable AI models is unlikely to curb global spending on AI infrastructure—If anything, it may accelerate it, according to Jefferies. In its new report, the brokerage cited the launch of Chinese AI company Z.ai’s GLM-5.2 model as “another DeepSeek moment“, arguing it intensifies competition among AI developers while sharply reducing the cost of inference—the process of running trained models to generate outputs from new data.

Many of these low-cost AI models like DeepSeek and GLM are open-weight versus the proprietary models. That said, open-weight models like Meta’s Llama and Mistral AI are not open-source since they only release the trained model, but not necessarily the training data, the training code, or the exact process used to build it. The term “weights” refers to the billions of numerical parameters that encode everything the model has learned during training.

According to OpenRouter’s recent analysis, The Open Weight Models that Matter, the performance gap between open-weight and proprietary (or closed) frontier models has remained remarkably stable at three to six months for more than 18 months. In other words, enterprises increasingly face a strategic choice—continue renting intelligence from a handful of Western AI providers, or own and operate their own.

That distinction matters

Unlike proprietary models accessed via cloud APIs, open-weight models can be downloaded, customised, deployed within an organisation’s own infrastructure. For many enterprises, this offers the prospect of lower costs, greater control, and reduced vendor dependence.

Yet, the rise of open-weight AI also exposes a growing confusion in discussions around sovereignty and security.

The fact that some of the world’s most capable open-weight models—such as DeepSeek, MiniMax and GLM (General Language Model)—originate in China does not mean that enterprise data automatically flows back to China. If an India-based bank downloads and runs such a model locally, the model creator does not receive the bank’s data. Nor does the phenomenon of “memorisation leakage”, whereby a fine-tuned model inadvertently reproduces sensitive information, constitute a violation of India’s data localisation rules in itself.

There’s another risk

Enterprises that access open-weight models through third-party cloud providers may expose their prompts and data to those providers. And as geopolitics increasingly shapes technology policy, questions of jurisdiction, auditability, and operational control may ultimately prove more important than the nationality of the model’s creator.

The future of AI competition, in other words, may hinge less on who builds the model than on who controls its deployment. In this context, India’s move to build sovereign multilingual, multimodal language models like Sarvam AI and BharatGen, makes a lot more sense.

AI BITS & BYTES

Google limits Meta’s use of its Gemini AI models

Google has reportedly put limits on Meta’s use of its Gemini AI models after the social media company sought more computing capacity than the rival technology group could provide. According to a Financial Times report, Google told Meta around March that it could not meet the full Gemini capacity the company had sought to purchase.

The report added that the shortfall disrupted and delayed some of Meta’s internal AI projects. Several other Google clients have also been affected, though to a lesser extent, according to the report. Meta has been particularly impacted due to its exceptionally high demand for Google’s models, the FT said.

Due to the restrictions, Meta has encouraged staff to be more efficient with AI tokens, the units that measure AI usage, the FT report said.

India’s robotics funding doubles but scale is off

Funding for India’s robotics startups almost doubled to $42.1 million in the first six months of 2026 from $22.7 million in the comparable span last year, with the average cheque size up 94.4% over the period, according to Tracxn data. The funding was $35.8 million in the first half of 2024.

While the rise in funding signals growing momentum in the sector, experts said it is still early days for India. In 2025, India’s robotics startups raised less than 1% of the capital secured by their US peers and 2% of the amount raised in China.

Apple prepares smart glasses to take on Meta

While the focus is on iPhone 18 Pro series and the upcoming foldable, Apple is slowly looking to enter new categories in order to shore up its revenue. As per a new report by Bloomberg’s Mark Gurman, Apple is looking to enter the smart glasses category currently dominated by Meta’s glasses.

Notably, Meta had just launched a cheaper version of its smart glasses last week with a starting price of $299. The glasses, which came in partnership with EssilorLuxottica, didn’t feature Ray-Ban or Oakley branding and feature a 12MP camera, 8 hours of battery life and dedicated speakers.

Meanwhile, Apple’s alternative to the Meta Glasses could be arriving sooner than most people would expect, with the Apple smart glasses tipped to arrive next year.

Google alerted users before Venezuela’s earthquake

Two powerful earthquakes that struck Venezuela on Wednesday evening have drawn attention not only for the widespread destruction they caused but also for the early warnings received by several Android users before the tremors were felt.

Social media users shared screenshots of Google notifications warning them of an approaching earthquake. One alert estimated a magnitude 6.2 quake about 212.3 miles (roughly 341 km) away, prompting many to ask how Google was able to send alerts before the ground started shaking.

The answer lies in Google’s Android Earthquake Alerts System, which turns millions of smartphones into a vast earthquake detection network.

Tech Talk is a weekly newsletter by Leslie D'Monte on everything happening in the world of technology and AI. Want this newsletter delivered to your inbox? Subscribe here.