BIG TECHAWAYS

WHAT MATTERS TODAY

​Anthropic locks up another $35 billion of AI compute​

The Wall Street Journal and Reuters report that Anthropic signed a $35 billion cloud contract with Lambda, an Nvidia-backed provider. The deal is reportedly tied to about 350 megawatts of Nvidia compute capacity at a Hut 8 data-center site in Texas. Anthropic, Lambda, Nvidia, and Hut 8 did not respond to Reuters outside business hours.

The $35 billion figure is the reported contract value, not recognized revenue, and 350 megawatts does not mean all of the infrastructure has been delivered. Although the parties have not directly announced the deal, its reported structure shows that AI labs cannot simply order GPUs. They also have to coordinate specialized cloud providers, financing, power, land, and data-center construction over several years.

Compute procurement is becoming part of model strategy. A lab that wants to train and serve frontier models must secure both capital and physical capacity before demand fully materializes. That pulls more companies into the same stack and makes model development increasingly dependent on long-term infrastructure contracts.

​Nvidia puts $3.5 billion behind a custom-chip ecosystem​

Nvidia announced a $3.5 billion investment in MediaTek convertible bonds and an expansion of the companies’ partnership. The new scope runs from custom XPUs and rack-scale AI systems to local AI computing and automotive platforms.

MediaTek already has a large presence in smartphones, connectivity, and edge devices. Combining that reach with NVLink Fusion gives Nvidia another route to place its interconnect technology, software, and rack architecture inside systems that do not necessarily use a standard GPU design. The custom-XPU partnership also lets hyperscalers and device makers build specialized chips while remaining connected to Nvidia’s platform.

The announcement does not prove that the resulting products will win in the market. It does show Nvidia extending its advantage beyond accelerator supply into the layers around the chip: interconnects, system design, software, and distribution to the edge. The AI hardware contest is increasingly about who can place an architecture across the widest range of systems, not only who can build the fastest accelerator.

PRESENTED BY WISPR FLOW

Talk to your AI tools the way you'd talk to a colleague.

You don't send a colleague a three-word brief. You explain the context, the constraints, what you've already tried. But typing all that into ChatGPT takes forever — so you don't.

​Wispr Flow lets you speak your prompts instead. Talk through your thinking naturally and get clean, paste-ready text. No filler words. No cleanup. Just detailed prompts that actually get you useful answers on the first try.

Millions of users worldwide. Works system-wide on Mac, Windows, and iPhone.

​Try Wispr Flow free​

BIG THINK

​Evaluate web search by the agent’s final answer​

Comparing search providers only by their raw results can produce the wrong conclusion. An agent’s answer also depends on its harness, model, search tool, extraction step, and synthesis. A provider with strong retrieval can still lose if the agent never calls the tool, writes a poor query, hits an infrastructure error, or mishandles information already in context.

Parallel Web Systems proposes the following process:

1. Choose the right task type. Search finds information, Extract reads a known URL, and a Task or Research API handles a broader objective and returns a sourced answer.
2. Build about 100 questions that resemble real production demand, including multi-hop, fresh, and domain-specific cases. Each question needs a verified gold answer.
3. Hold the agent harness, model, prompt, budget, and judge constant. Change only the search tool, and allow the agent multiple turns within the same total search budget.
4. Grade the final answer and its citations. Depending on the task, measure correctness, recall, precision, field-level accuracy, or use a rubric for open-ended research.
5. Manually inspect at least 10% of both successes and failures. Separate errors into four groups: the agent did not search, the provider failed, retrieval missed the information, or synthesis failed.
6. Measure cost per resolved task, total tool calls, token usage, and p50, p95, and p99 latency. Compare accuracy against cost and latency, add confidence intervals, and treat overlapping results as ties.

Parallel is a search provider, so product-specific configurations and performance improvements come from the company itself. The evaluation process has broader value: keep the experiment fair and grade the entire system by the agent’s final output.

SHIFT SIGNALS

BEHIND THE HEADLINES

• ​OpenClaw 2.0 moves agent operations into the product core: OpenClaw 2.0 brings durable progress, scoped permissions, automation, memory, recovery, and interoperability into one operating layer. Preserving state, resuming work after interruptions, and handing tasks between people and agents are becoming core product capabilities, rather than relying only on the quality of each model call.
• ​Runway uses a world model to render interfaces in real time: Runway introduced Solaris, a world model that generates an interactive interface frame by frame instead of rendering a UI from traditional code. The company says the system can respond to interactions in real time. Its performance comparisons remain company-reported, but the demo points toward generative software that creates the interactive experience itself, not just the underlying code.
• ​TimesFM-3 forecasts multiple variables together: Google Research introduced TimesFM-3, a 330-million-parameter model pretrained on more than one trillion time points. It supports multiple targets and covariates for zero-shot forecasting. The benchmark results and outperformance claims currently come from Google’s own research team.
• ​ChatGPT Ads reaches a $1 billion annualized revenue run rate: OpenAI says ChatGPT Ads has reached a $1 billion annualized revenue run rate as self-serve buying expands into more markets. This is not audited full-year revenue, but advertising is becoming a meaningful revenue stream alongside subscriptions and APIs.

PRESENTED BY EZ TEXTING

Your Customers Are One Text Away

You already know email gets buried and calls go unanswered. But texts? They get read — 98% of the time, within minutes. That's why 230,000+ businesses use EZ Texting to reach their customers on the channel they engage with most Whether you're sending promos, appointment reminders, or urgent alerts, EZ Texting makes it easy to send to 10 or 10,000 contacts in just a few clicks. No tech skills required. Techaways readers can try it free — no credit card needed.

​Start Your Free Trial with EZTexting​

Offer valid for new customers only. Use promo code EZWORKFLOW at checkout.