DX Today AI Daily Brief - Friday, August 14, 2026

DX Today AI Daily Brief - Friday, August 14, 2026

Google introduces Gemini 3.7 Flash, a coding and agent model priced at 75 cents per million input tokens, while its top-tier Gemini model stays delayed. DeepSeek moves V4 Pro to general availability with roughly 1.6 trillion parameters and a one million token context window under an MIT license, and Z.ai ships GLM-5.3 without retraining its base model, with weights promised in two weeks. IBM and OpenAI announce an enterprise partnership embedding GPT-5.6 in IBM Consulting Advantage, and Databricks closes a $5 billion round at a $190 billion valuation. Anthropic is in advanced talks to buy Israeli startup Decart for about $6 billion, which would be its largest acquisition ever. Apple trains its own China-market model with Alibaba's support. Applied Materials posts record fiscal third quarter revenue near $9.1 billion and still sees its stock fall, while Lenovo reports the best quarter in its history at $26.9 billion. AMD raises up to $5 billion in bonds to fund AI expansion, Jeff Dean's new venture Discovery Loop seeks $1 billion at a $10 billion valuation, and SK Hynix details a roughly $720 billion memory buildout in South Korea.

DX Today - Your daily AI news briefing, delivered every morning.

#AINews #ArtificialIntelligence #ModelLaunch #Semiconductors #DXToday

Transcript

SPEAKER_05 0:00

It's Friday, August 14th, 2026. You're listening to the DX Today AI Daily Brief. Today, Google ships a new Gemini model while its flagship stays delayed. Two Chinese labs push open weight coding systems toward the frontier within hours of each other. And Anthropic weighs the largest acquisition in its history. Let's get into it.

SPEAKER_00 0:24

Google introduced Gemini 3.7 Flash on Thursday, a coding and agent focused model that arrives roughly three weeks after the previous Flash release. The company priced it at $75 per million input tokens and $3.75 per million output tokens, positioning speed and cost as the pitch rather than raw capability. Google says the model is tuned for long-running agent workflows and software development, the same territory where rivals have concentrated their recent releases. The launch lands with an awkward asterisk. Google's top-tier Gemini model remains delayed, and the company has not offered a firm date. For now, the cheaper, faster tier is doing the talking.

SPEAKER_02 1:08

From Mountain View to Hangzhou, DeepSeek formally released its V4 Pro model on Thursday, moving the system out of preview and into general availability across its app, its web interface, and its API. The model carries roughly 1.6 trillion total parameters in a mixture of experts design with about 49 billion active per token, and ships with a 1 million token context window under an MIT license. DeepSeek is emphasizing agent capability, the ability to plan and execute multi-step tasks rather than simply answer questions. The release is the clearest signal yet that the Chinese lab intends to compete at the frontier on open weights rather than retreat behind a paywall.

SPEAKER_05 2:02

A second Chinese release, hours later.

SPEAKER_04 2:05

Z.ai, the company formerly known as JIPU, launched GLM 5.3 today. A coding-focused model, the lab says it built without retraining its base model at all. Instead of a fresh pre-training run, Z.ai leaned entirely on post-training scaling, and it is publishing benchmarks that place the model above Anthropics Claude Opus 4.8 on complex coding and long horizon agent tasks. The weights are not public yet. The company says it will release them in about two weeks, once safety evaluations conclude. That gap is worth watching because open wait timing has become its own competitive lever in a market where every lab is racing the same calendar. Now to enterprise deployment.

SPEAKER_01 2:50

IBM and OpenAI announced a strategic enterprise partnership on Thursday that puts OpenAI's frontier models, including GPT 5.6, directly inside IBM Consulting Advantage. The agreement covers core operations, systems modernization, and cybersecurity work, and it elevates IBM into OpenAI's elite partner tier. IBM is also standing up a dedicated OpenAI practice staffed by thousands of certified consultants and engineers. The strategic logic is straightforward. Whether the consulting margin survives contact with the model economics is the open question.

SPEAKER_05 3:39

Capital keeps flowing.

SPEAKER_03 3:41

Databricks closed a $5 billion strategic funding round on Thursday at a $190 billion valuation, among the largest private raises of the year. The company says the capital goes towards scaling its enterprise AI platform, naming Lakebase, Genie, and its Unity AI Gateway as the priorities. That product list matters more than the headline number. Databricks is betting that enterprises will standardize on a governance and data layer beneath whichever models they choose, rather than committing to a single model vendor. The valuation implies investors are buying that thesis. It also sets a marker for a company widely expected to test the public markets.

SPEAKER_05 4:25

And a deal in the making.

SPEAKER_00 4:26

Anthropic is in advanced talks to acquire the Israeli startup Descartes for about $6 billion, according to Bloomerberg. If it closes, it would be the largest acquisition in Anthropic's history by a wide margin. Descartes works on real-time video generation and chip level optimization, capabilities that sit closer to infrastructure than to the assistant products Anthropic is known for. The talks are not final and could still collapse, but the reporting arrives as Anthropic is widely believed to be preparing for a public offering, which makes a purchase of this size a statement about what the company thinks it needs to own outright rather than rent.

SPEAKER_02 5:05

Meanwhile, in Cupertino, Apple has trained its own large language model specifically for the China market, with technical and cloud support from Alibaba Reuters reported today, citing sources. The model is intended to underpin Apple intelligence inside China, where regulatory requirements have kept the feature set largely unavailable. It marks a shift away from relying solely on third-party Chinese models, though the picture is more layered than a clean break. Apple still integrates Alibaba's Quen and Baidu's models within the China-specific stack. What has changed is that Apple now owns a piece of the model layer itself, in the one market where it has been most conspicuously behind.

SPEAKER_04 6:00

Thursday, with record revenue of about $9.1 billion, up roughly 25% year over year, beating consensus and prompting the company to raise its outlook. Management credited AI-driven semiconductor demand, particularly in advanced logic and high bandwidth memory tooling. Investors were unimpressed. The stock fell close to 5% in after hours trading. That reaction has become a pattern across this earnings season. Beating expectations is no longer enough when expectations already assume a permanent buildup, and the market has started scrutinizing the shape of the guidance rather than the size of the beat. Another record, a different reaction.

SPEAKER_01 6:43

Lenovo posted what the company called the best quarter in its history, with record revenue of $26.9 billion and record-adjusted net income of just over $1 billion. Growth came across all three business groups rather than from a single AI server line, which is the detail that separates this result from most of its peers. The infrastructure solutions unit benefited from AI server demand, but the personal computing business grew as well, suggesting the refresh cycle analysts have been forecasting for two years may finally be arriving. Lenovo has spent that time repositioning from a hardware assembler toward a services and solutions mix.

SPEAKER_05 7:25

Financing the build-out.

SPEAKER_03 7:26

AMD moved to raise as much as $5 billion in a four-part investment grade bond sale on Thursday, with proceeds earmarked for AI expansion, general corporate purposes, and potential debt repayment. Reporting indicates the offering ultimately priced at about $4.75 billion. The raise sits alongside AMD's large compute commitment to Anthropic, and it fits a pattern now visible across the sector. Chip companies are increasingly funding capacity with debt rather than cash flow because the build-out schedule has outrun the revenue schedule. That is not inherently alarming for a company with AMD's balance sheet, but it is a change in posture worth noting.

SPEAKER_05 8:10

A new venture takes shape.

SPEAKER_00 8:12

Jeff Dean, the longtime Google chief scientist whose departure was confirmed earlier this month, is in talks to raise about $1 billion for a new startup called Discovery Loop at a valuation near $10 billion. That is according to multiple reports, citing people familiar with the discussions. Neither Dean nor the company has confirmed the figures. The valuation is remarkable for an entity with no announced product, and it reflects a market where a founder's research pedigree is priced as an asset in its own right. Dean spent more than two decades at Google building the infrastructure that made modern deep learning practical.

SPEAKER_02 8:50

And finally, memory. CNBC published a detailed look Thursday at S.K. Heinrich's roughly $720 billion bet on memory capacity for AI, centered on its Yongen and Chongju clusters in South Korea. The figure is a long-term multi-tranch plan running to about 2034, not a single year of spending, and portions of it are denominated in one before conversion. The scale still reframes the constraint. High bandwidth memory, not logic, has become the tightest link in the AI supply chain. And SK Heinrich is effectively wagering that the shortage lasts long enough to justify a decade of construction.