Neural Dispatch: Tense times for Nvidia’s world view, and AI refuses to take responsibility (HT Tech)

0
11
Neural Dispatch: Tense times for Nvidia’s world view, and AI refuses to take responsibility (HT Tech)


Cognitive Warmup. RAM prices are rising, and it won’t be good. Mark my words. The fact that 96GB DDR5 RAM costs $900 (approximately) 89,000) means it’s more expensive than an entire Sony PlayStation 5 (approx. 54,990). These are memory prices that have not doubled, but tripled in the last few months. The main culprit? AI’s insatiable appetite for memory. Data centers and cloud providers are purchasing huge amounts of high-bandwidth memory for AI servers. The consumer RAM business naturally takes a back seat – that means your next smartphone, desktop, laptop, even your next tablet or gaming console is going to be a little more expensive. And I am not exaggerating the scenario. There is another element at work in this. The transition from DDR4 to DDR5 (the first DDR5 products began arriving in 2022) increased complexity, and initial production of the new nodes markedly reduced yields, leading to supply constraints during this crossover period between 2021 and 2023. Memory brands including Micron, Crucial, Patriot and Corsair were playing inventory catch-up. Unfortunately, relief is not coming soon. I don’t expect prices to drop meaningfully before mid-2026. Meanwhile, tech companies will essentially pass these costs on to customers. This doesn’t mean you should rush out to buy a new phone or laptop, even if you don’t need it.

Google, which has largely rented out custom TPUs until now, is offering them for sale for the first time

Last time on Neural Dispatch: AI math is broken, and a bubble that many pretend not to see is bursting

algorithm

This week, we talk about stressful times for Nvidia, as Google makes a big move for its TPU business that directly hurts Nvidia’s GPU sales, and it’s a case of never say never as OpenAI admits it can be a matter of when and if not to advertise within ChatGPT.

Google TPU vs Nvidia GPU: Who’s Winning!

“Competitors could price their chips at $0 and Nvidia products would still be a better choice,” CEO Jensen Huang said recently. This didn’t make aging any better at all, did it? Meta is now buying billions of dollars worth of Google TPUs, or Tensor Processing Units (not immediately, but in a gradual manner). There are two important things to note here. First, Google, which has largely rented out custom TPUs until now, is offering them for sale for the first time. Second, despite Nvidia’s insistence to the contrary, TPUs as application-specific integrated circuits (ASICs) would be a better and more cost-effective fit into typical AI frameworks. This news had a deep impact on Nvidia’s stock price. So much so, that Nvidia posted a disappointing congratulatory note (after all, everyone still has to pretend they’re friends, for AI’s sake…) claiming that “Nvidia is a generation ahead of the industry – it’s the only platform that runs every AI model and does so everywhere computing is done.” Great.

What’s the difference between the GPU that Nvidia wants to sell for AI tasks, versus the TPU that Google champions? GPU is a specialized processor originally designed to manipulate computer graphics. Their parallel structure makes them ideal for algorithms that process large blocks of data typically found in AI workloads. TPUs are an ASIC designed with special features for neural networks, such as matrix multiple units (MXUs) and proprietary interconnect topologies that make them ideal for accelerating AI training and inference. All Google AI services run on TPU. Right now, Google has three iterations available to customers – Cloud TPU v5e which is ideal for medium to large-scale training and inference workloads, Cloud TPU v5p which is a powerful TPU for building large, complex fundamental models, as well as Trillium, a sixth-generation TPU that has pushed the standards for energy efficiency and extreme compute performance per chip for training and inference. Next in line is Ironwood TPU, which Google says is “the most powerful and efficient TPU ever for large-scale training and inference.”

One thing is clear. Suddenly, the ‘Nvidia or nothing’ worldview seems less absolute. Is Nvidia still the gold standard? Perhaps, but we’ll let market forces decide. But the armor no longer shines.

Nvidia says they’re not like Enron

I mean, I don’t know everything, but I would say that if you have to say, “We’re not like Enron,” then you’re already on the back foot. That said, Nvidia’s efforts to calm growing speculation over revenue recognition, accounting methodology and the realistic lifespan of GPUs are looking like a company that is on the backfoot. In a note to Wall Street analysts in response to the AI ​​bubble certainly being punctured over the past few months (causing panic as circular funding gas is revealed) and US investor and hedge fund manager Michael Burry also made a clear post X on Nvidia. earnings releaseHe pointed to the practice of Nvidia customers now using a six-year depreciation schedule instead of the standard two- or three-year practice, Bury argued about the mathematics of that practice, Still, Nvidia said in the note, “Nvidia does not engage in historical accounting fraud because Nvidia’s underlying business is financially strong, our reporting is complete and transparent, and we care about our reputation for integrity, Unlike Enron, Nvidia does not use special purpose entities to hide debt and inflate revenues,” great then, The fact that investor forums, analysts, and even some very vigorous short-sellers managed to find enough shade for that comparison is telling, As every over-performing tech giant eventually learns – gravity isn’t a rumor; This is a schedule,

From “never advertising”, it…

It was inevitable, right? It looks like the world’s most powerful conversational product is making a move towards the world’s oldest monetization model – advertising. Some things never change, no matter how much technology changes. OpenAI’s Sam Altman indicated that advertising is “something we’ll try”. Don’t be put off by this, because in OpenAI-speak, this translates to “There’s a pilot already running somewhere, we just haven’t told you yet.” I guess that means three things. First, AI search is going to look even more like traditional search. Secondly, “objective answers” may gradually become “sponsored objectivity”. And third, any dream of an ad-free, neutral, logic-first AI assistant can go to the retirement home. Ray of hope? If done with transparency, advertising can subsidize specifically India-centric education initiatives at affordable levels, without diluting the integrity of the product. But the Internet has heard that promise before. And some of us will only believe what we see.

Be sure to check out my other newsletter, Wired Wisdom: Decoding the home air purifier, SaveSage’s AI saves you money, and Toyota’s temple museum

Thinking

“We will respectfully present our case in a way that is informed by the complexity and nuance of real people and real-life situations… Because we are the defendants in this case, we are required to respond to the specific and serious allegations in the lawsuit.” – OpenAI’s response to a lawsuit, and denial of liability in teen suicide.

While this may be another corporate statement crafted by a legal team to protect larger business interests (as far as possible), it signals an uncomfortable inflection point. While AI systems have become so emotionally accessible that people trust them just as they do humans – they are still engineered to behave like tools. No limit. Asim. no balance. I am reminded of the time when Meta CEO Mark Zuckerberg was solemnly admonished by the US Senate Judiciary Committee during a hearing on online child protection and exploitation (it was on January 31, 2024), and was forced to apologize to parents of children who were harmed by interactions on social media. Technology never had social filters for interacting with children. This is at a time when the mismatch of public expectations is increasing. Users increasingly (and quite unfairly, I might add) regard ChatGPT as a considerate, empathetic entity. It is not, and there is not and never will be any AI. Regulators still classify it as software, and rightly so. It is a tool and always will be – tools cannot be human, humans have to use the tools properly. Now is the time when courts must interpret a technology whose effects are human-like, yet whose accountability is legally non-human. This reaction is an example of that.

A reality check: The lawsuit I’m referring to is a manifestation of that cognitive dissonance. And it won’t be the last. As AI becomes more agentic and reportedly more personalized, finding itself more deeply embedded in mental-health-adjacent interactions, the boundary between impact and responsibility is becoming blurred still.

OpenAI’s defense, that the company cannot be held liable for all user outcomes, is institutionally predictable. No AI firm can survive if it is held responsible for every emotional, psychological or behavioral outcome arising from a model’s output. But morality does not take its dictates from legal disclaimers. And the truth is – AI companies have spent years marketing these systems as empathetic, conversational, human-like companions. You can’t sell “intelligence,” “logic,” and “connections” on Monday and then argue “we’re just a tool” on Wednesday afternoon without raising eyebrows and lawsuits. The question is not just what happened in this tragic case. They are about who bears the burden when AI steps into human emotional territory. experimenter? developer? regulator? Or is no one prepared to define the emerging gray zone of shared responsibility?

The legal phase of the AI ​​era has officially begun, and the industry can now (and very uncomfortably, I might add) begin to discover that being “transformational” also means being accountable.


LEAVE A REPLY

Please enter your comment!
Please enter your name here