Tech Brew Ride Home - Is Nvidia In Trouble?

Episode Date: December 1, 2025

While Runway releases a new video model, let me break down the big analysis piece that had everyone concern trolling about Nvidia over the weekend. Why doesn’t Netflix want you to cast to your tv anymore? And AI means less jobs in consulting, but more jobs in a specific type of construction. Runway rolls out new AI video model that beats Google, OpenAI in key benchmark (CNBC) TPUv7: Google Takes a Swing at the King (SemiAnalysis) Nvidia takes $2 billion stake in Synopsys with expanded computing power partnership (CNBC) OpenAI Takes Stake in Thrive Holdings, a Buyer of Services Firms (NYTimes) Netflix kills casting from phones (The Verge) Top consultancies freeze starting salaries as AI threatens ‘pyramid’ model (FT) Data Centers Are a ‘Gold Rush’ for Construction Workers (WSJ) Learn more about your ad choices. Visit megaphone.fm/adchoices

It's peak pollination season, and my business is scaling fast. To keep the nectar flowing, I need a phone plan with top priority data speed. That's why I chose GoogleFi Wireless. My connections stay strong even when the hive is buzzing. Plus, unlimited plans start at $35 a month. Now, that's a deal that doesn't stay. Explore GoogleFi Wireless plans today. Plus taxes and government fees.

GoogleFi Wireless is not subject to data traffic deprioritization during times of high network usage. Welcome to the Tech Brew Ride Home for Monday, December 1st, 2025. I'm Brian McCullough today. While Runway releases a new video model, let me break down the big analysis piece that had everyone concern trolling over Invidia over the weekend. Why doesn't Netflix want you to cast to your TV anymore? And AI means less jobs in consulting, but more jobs than a specific type of construction. Here's what you miss today in the world of tech. Conducting business online can feel a little scary these days, especially with AI creating new opportunities for fraud. In fact, Sellant estimates that AI was behind roughly 20% of the fraud perpetrated in 2024.

Spotting bad agentic AI while allowing good agents to continue with their tasks isn't easy. Thankfully, Momoto Continuous Captcha can spot malicious agents pretending to be people at the point of account creation or registration. Unlike past Captcha solutions, it runs behind the scenes with no puzzles for users. Momoto offers TechBrew Ride Home listeners early access with a special price for Momoto Continuous Captcha. Right now, our listeners can purchase a year of Momoto Continuous Captcha for $5,000, a 20% discount on their lowest price plan. To learn more, head to mimoto.com. A.I. slash ride home. That's mimoto.com. A.mode slash ride home. Well, we're back at it, and they're back at it with another new model runway. has launched Gen 4.5, a new text-to-video AI model that produces HD videos from written prompts

and excels at physics. Gen 4.5 already apparently tops Video Arena's leaderboard, quoting CNBC. Google's V-O-3 model holds second place on that leaderboard, and OpenAI's SORA2 Pro model is in seventh place. We managed to out-compete trillion-dollar companies with a team of a hundred people runway CEO Cristobal Valenzuela told CNBC in an interview, you can get to Frontiers just by being extremely focused and diligent. Valenzuela said Gen 4.5 was codenamed David in a nod to the biblical story of David and Goliath. The model was, quote, an overnight success that took like seven years, he said. It does feel like a very interesting moment in time where the era of efficiency and research is upon us, Valenzuela said, we're excited to be able to make sure that AI is not

monopolized by two or three companies. Gen 4.5 is rolling out gradually, but it will be available to all of Runway's customers by the end of the week. Valenzuela said it's the first of several major releases that the company has in store. It will be available through Runway's platform, its application programming interface, and through some of the company's partners, he said, end quote. Last week when I did that piece about Google may be shopping its TPUs around to outside companies as direct competition to Nvidia's GPUs, I remember thinking, why aren't people talking about this? Why isn't this a bigger deal? Well, over the last week, it's become a big deal weighing heavily on Nvidia's stock. And over the holiday weekend, there was a lot of chatter about

this piece from semi-analysis breaking down why people think this is potentially a big deal. The piece is long, so I'll do my best to summarize it for you, but I've also linked to it in the show notes so you can read the whole thing for yourself. Basically, semi-analysis argues that Google's latest AI chip system, TPUV7 Ironwood, plus a big shift in Google's business model, is the first serious structural challenge to Nvidia's dominance in high-end AI hardware. They point out that in AI right now, hardware dominates costs. Buying and running chips is a much bigger line item than paying programmers, for example. That's a big change from the recent software is eating the world decade. Because that means whoever can train and run models on cheaper, more efficient

infrastructure has a real business advantage. Historically, NVIDIA, of course, has owned this space with its GPUs and Kuda software ecosystem. Now, two of the strongest AI models out there, Anthropics Clod 4.5 Opus and Google's Gemini 3, run mostly on non-invideo hardware, Google's TPUs and Amazon's Traneum. Google originally built TPUs as internal AI-specific chips to avoid doubling its data center footprint when deep learning took off around 2013. For years, they used it largely to power search and ads for them.

And for years, they were mostly an internal advantage with only half-hearted cloud offerings to others. That is now changing. Google is turning TPUs into true merchant silicon that it sells or rents at scale, not only via Google Cloud platform, but also as physical systems to third parties, most notably anthropic,

and likely potentially down the road meta XAI, OpenAI, and others. Those other deals may or may not happen, but the piece takes a deep look at Anthropics' big bet on TPUs recently as a way to cut costs drastically internally for them. They've contracted 1 million TPUV-7 chips, roughly 400,000 TPUs. About $10 billion worth of hardware will be bought directly from Broadcom, which manufactures the TPUs and co-designs them with Google to install those in Anthropics' own facilities. The remaining around 600,000 TPUs will be rented from Google Cloud. Semi-analysis estimates this rental piece alone represents around $42 billion of backlog for Google

and accounts for most of the recent $49 billion increase in GCP's reported remaining performance obligations. By itself, the very idea of external competition already weakens Nvidia's pricing power. Even before OpenAI actually started using TPUs itself, the credible threat of moving some of its workload off of Nvidia, let OpenAI negotiate roughly 30% savings on its Nvidia fleet. The author's point is that TPUs can cut Nvidia's capital expenditure even when no TPU is powered on simply by existing as a realistic alternative. Competition is a hell of a thing, but there's more because just on raw specs, while previous TPU generations lagged NVIDIA on advertised peak, compute, and memory. Google was prioritizing reliability and its own recommendation system workloads over

headline flop numbers. Now, with the LLM wave, that philosophy has shifted for Google. TPUV6 and now TPUV7 move much closer to NVIDIA's current GB200 and GB300 GPU systems in peak compute and bandwidth, while still shipping slightly later. TPUV7 uses high-end HBM 3E memory and nearly matches Nvidia's top chips in theoretical performance. But the authors stress that what matters is effective performance per total cost of ownership, TCO, not the headline flop numbers. They argue that Nvidia and AMD inflate peak flop figures using clock speeds and test conditions that are not sustainable in real workloads, realized utilization can be a fraction of the theoretical maximum. Google's TPUs, by contrast, have more conservative.

internally driven specs, and when paired with Google's system-level engineering and good compiler work can reach higher-effective utilization. Combined with procurement economics, NVIDIA takes very high margins on the full system. Google does have to pay Broadcom, who in turn makes margin, but that is still less than NVIDIA's stacked margin. TPUV-7's full system total cost of ownership is estimated to be around 44% lower than a GB-200 server for Google itself, and still roughly 30 to 40, percent cheaper per hour for large external buyers like Anthropic. That translates into more than 50 percent lower cost per effective training flop for Anthropic versus Nvidia's GB300-based systems, even if TPUs run at somewhat lower utilization. The article then zooms out to Google's networking design.

TPUV-7 systems are built as 64-chip 3D cubes wired into a tourist pattern within a rack, and many such cubes can be connected with optical circuit switches into very large world sizes, up to 9,216 TPUs tightly coupled for training, and up to about 147,000 TPUs on a broader data center network. This scale is far beyond the 64 to 72 GPU pods common in the rest of the market. Optical switching also lets Google reconfigure clusters flexibly, route around failures, and expand capacity without tearing up the entire network. The result is lower latency, better locality, and lower networking costs per unit of compute at very large scale. A separate section describes how this push reshapes the neocloud market.

Smaller cloud providers and repurpose Bitcoin miners that rent out GPU clusters. Data Center power is now the bottleneck and crypto miners already control power contracts and electrical infrastructure. Google's workaround for its own slow contracting process is to give these neoclouds an off-balance sheet, credit backstop, essentially an IOU that Google will take over capacity if the Neocloud can't pay its lease. That template reduces financial risk for both data center builders and Neocloud and is likely to spread throughout the industry, creating a path for many more crypto-to-a-conversions. The last big theme is software. Everybody knows that NVIDIA's real estate moat is Kuda, and the vast ecosystem of libraries and tools therein, not just the chips. Historically,

Google's TPU software focused on internal frameworks like Jax and TensorFlow, external developers had a clumsy pie torch on TPU experience with non-standard APIs, which limited adoption. Google is now shifting hard, however. It is putting major engineering effort into making TPUs a first-class pie torch backend with proper eager execution and native distributed APIs, and into integrating TPUs with widely used inference stacks like VLLM and SG Lang. It is also open sourcing optimized TPU kernels for attention, mixture of experts, etc. and investing in kernel languages like Pallas and related compiler work to make it easier for others to reach high utilization. The bottom line from this piece, TPUV7 already offers roughly similar raw performance to NVIDIA's leading hardware,

materially better economics for very large buyers and unique system-level advantages in networking and scale. If Google continues to improve and opens up the TPU software stack, especially its compilers and large cluster management tools, the Kuda Mote could narrow, forcing NVIDIA to compete more on price and reducing its ability to extract outsized margins from the AI boom. NVIDIA's fat margins could, very much, be Google's opportunity. Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the unreal college deal, everything you need, to study and play with select Windows 11 PCs.

students get a year of Microsoft 365 premium and a year of Xbox GamePass Ultimate with a custom color Xbox wireless controller. Learn more at Windows.com slash student offer. While supplies last, ends June 30th, terms at AKA.m.m.m.a. Resort and Casino at San Manuel is California's number one entertainment destination for today's superstars. Catch the Jonas Brothers return to the Yamava Theater stage on April 30th, the powerful vocals of Demi Lovato on May 17th, and the signature Southern Country Rock of Eric Church on July 19th. Tickets on sale now at Yamava Theater.com, only at Yamava Resort and Casino, celebrating its 40th anniversary. You win? Must be 21 to enter.

Ready to soundtrack your summer? With Red Bull Summer All Day Play, you choose a playlist that fits your summer vibe the best. Are you a festival fanatic, a deep end DJ, a road dog, or a trail mixer? Just add a song to your chosen playlist and put your summer on track. Red Bull Summer All Day Play. Red Bull gives you wings. Visit Red Bull.com slash bright summer ahead to learn more. See you this summer.

Now, all of that ties into the whole idea that both Nvidia and OpenAI knew that their positions as the AI leaders would come under threat eventually. Thus, the cynical thinking goes, they've been pursuing these so-called circular deals to sort of build a too big to fail moat around themselves. Throw your money around, investing in your customers, they can continue to be your customers. Again, that's the cynical view, not necessarily one I subscribe to, but, well, Nvidia has acquired $2 billion worth of synopsis' common stock and has unveiled

a strategic partnership with them to accelerate computing and AI engineering products, including deploying Kuta. Quoting CNBC, Synopsis offers services like Silicon Design and electronic design automation that helps its customers build AI-powered products. CEO Sassine Gazi told CNBC that its partnership with NVIDIA will help it take workloads that used to run for weeks and reduce them to hours. We're going through a platform shift from classical general purpose computing running on CPUs to a new way of doing computing, accelerated computing running on GPUs. NVIDIA's Jensen Wong said that old way of doing is going to continue to exist, of course, but the world is shifting to this new way of doing computing, end quote.

And V-Vyna and Synopsis have a long-standing relationship, so Monday's announcement builds on their existing partnerships, end quote. And Open AI has taken a stake in Thrive Capital's Thrive Holdings and plans to embed AI agents in the companies in that portfolio, which already includes an accounting and IT business, quoting the times. For OpenAI, the hope is that working more closely with Thrive Holdings will demonstrate how companies can harness the technology behind OpenAI's chatGPT chatbot and create buzz to entice other potential customers. What we're trying to do with this partnership is really prove out ways that we can accelerate that type of transformation, Brad Lightcap, OpenAI's chief operating officer said in an interview. Thrive Capital created Thrive Holdings this year with an initial $1 billion in funding. The vehicle aims to do serial dealmaking known in financial lingo as roll-ups in relatively humdrum industries that it says would benefit from AI, a strategy that other venture capital firms have

embarked on. Thrive Holdings two current operations, the accounting business Crete Professionals Alliance and the IT services provider shield technology partners have more than 1,000 employees in total. Thrive Holdings has committed $500 million to Crete, which the trade publication accounting today described this year as one of the fastest growing accounting firms in the United States. Crete, for example, has been working to use the technology to automate tasks such as data entry and processing tax returns to help free up accountants to work more directly with clients, and quote. Netflix has quietly killed support for casting from its mobile app to most modern TVs and streaming

devices, including Chromecasts, regardless of your subscription tier. Quoting The Verge, Netflix has removed the ability to cast shows and movies from phones to TVs unless subscribers are using older casting devices. An updated help page on Netflix's website, first reported by Android Authority, says that the streaming service no longer supports casting shows from a mobile device to most TVs and TV streaming devices and instead directs users to navigate Netflix using the remote that came with their TV hardware. The change seems to have rolled out in the last few weeks with one user on Reddit reporting that casting support was removed on November 10th with zero warning. My colleague

Dom Preston also found that while he was able to cast to a TV from an older version of the Netflix app, the casting option was no longer available after the app was updated. Casting support is still available on older Chromecast devices or TVs that support Googlecast natively, according to Netflix's support page, but only for subscribers on pricier ad-free plans, which start from $1799 per month. Netflix users with an ad-supported subscription at $7.99 per month will be unable to cast from their phones even if they own legacy Chromecast devices. The casting change is announced on Netflix's support page, do not explain why the feature has

been removed. It follows a similar move in 2019 when Netflix removed Airplay support, citing a desire to, quote, ensure our standard of quality for viewing is being met. We have reached out to Netflix for comment, end quote. Finally today, two contrasting stories of the AI impact on jobs. First, McKinsey and other top consultancies have apparently frozen graduate pay offers in 26, making this the third straight year they've done so as AI is reshaping the industry and threatens its so-called pyramid model, quoting the FTE. Firms were seeking more mid-career specialized staff as they spent last on traditional strategy consulting and more helping companies implement technology and AI, said a representative

for McKinsey, it is harder to staff a 23-year-old on those kinds of projects versus someone with experience, the spokesperson said. Some industry executives say that the more conservative hiring practices are in anticipation of productivity gains from AI, not because those gains are necessarily being realized yet. Two senior executives at Big Four firms estimated that across the UK's largest consulting and accounting firm's graduate recruitment would be down by about half in the coming year. Some of that is commercial because the market's tougher, but some of that is anticipation of the impact of AI. Employment costs are going up because of national insurance, the minimum wage, etc. And you might be in a better place investing in AI and offshoring than in people, one of the

executives said. The upheaval means that the traditional pyramid in which a firm employs thousands of junior-level employees and thins out the ranks with an up-or-out promotion culture could be set to change, according to experts. Some are betting on an obelisk structure with fewer layers and less reliance on junior staff, while others predict an hourglass, pinched in the middle as AI automates mid-level routine tasks, end quote. Meanwhile, conversely, this is from the journal. The AI boom has led to high demand and more pay for the construction workers that build data centers. A trade group says that there is a shortage of around 439,000 workers in North America. Quote, data centers don't employ many workers once they are actually built, but during construction, they are a hive of

workers pouring concrete walls and foundations, wiring electric panels, and installing equipment such as power generators and chillers to ensure servers are cooled to a precise temperature at all times. Given such complexity and high demand, workers who move into the data center industry in roles ranging from electricians to project managers often earn 25 to 30 percent more than they did before, said Jake Razweiler, senior vice president of data centers at Kelly services, a staffing and recruitment firm. It's like the Gold Rush, Chambliss said. In Hermiston, Oregon, Mark Benner, 60, arrives in the pre-dawn hours at a data center construction site and lines up with scores of workers for a series of synchronized stretches. After that, he spends the day making the rounds

ensuring electrical safety. These are lucrative skills at the electricity gobbling sites, and Benner makes $225,000 a year boosted in part by $100 in daily incentive pay for all workers on site. It's my American dream, said Benner, who has been helping build data centers for 15 years, including the ones now powering AI. Demand for such workers is colliding with a longstanding shortage of skilled tradespeople that has pinched the construction industry. The Associated Builders and Contractors Trade Group estimates that the construction industry is short roughly 439,000 workers, mostly among skilled workers who do things like lay

pipe and wire electrical panels. The effects are starting to pile up. A survey by the Uptime Institute of Data Center equipment manufacturers, engineers, and construction companies found that 52% said staffing shortages on sites had caused business disruptions up from 43% last year. Contractors working on data centers have an average backlog of 10.9 months of work compared with eight months for their peers, according to data from ABC, end quote. Nothing more for you today. Talk to you tomorrow.