SAP Halts Travel and Hiring Over High AI Costs, AMD Bolsters Its Chip Lineup With Taalas Acquisition, Humans Miss a Third of Dangerous Agent Requests — AI News for August 7

SAP Halts Travel and Hiring as AI Costs Soar, Bringing Enterprise Spending Pressure Into Sharp Relief

German software giant SAP has reportedly suspended most business travel and hiring in response to the skyrocketing cost of AI (artificial intelligence) usage. The move is drawing attention as a concrete example of how the burden of AI investment has begun to squeeze existing business operations themselves.

On the question of "how AI will change the software business," ITmedia published an article examining the "death of SaaS" thesis. The author, a product manager with 40 years of experience, revisits — through the history of business software — whether the widely cited scenario, in which the spread of AI agents shrinks the SaaS user base and with it vendor revenues, actually holds. Far from mere pessimism, the piece takes a fresh look at the value these tools deliver to enterprise users.

The impact on the labor market is also being reported. An analysis by Bank Underground (a staff blog of the Bank of England) points to a slowdown in UK hiring in occupations with high AI exposure, describing it as data worth watching as "a harbinger of the AI wave." From three angles — cost, business models, and employment — AI has started to shake the foundations of business and labor.

Humans Miss "One-Third" of Dangerous AI Agent Requests

On the safety of AI coding agents, a research finding that warrants caution has been reported. According to The Register, human reviewers miss roughly one-third of the requests issued by dangerous AI coding agents. The result exposes just how fragile the practice of keeping "a human in the loop" to oversee agent decisions can be — more so than expected.

Barriers to adoption have also been pointed out. WIRED, in a piece titled "Why Consumers Aren't Using AI Agents," reports that the tech industry is beginning to realize that agents need to be designed around what consumers actually want, not around model capabilities. Even as performance improves, the piece argues, the experience design needed to blend agents into people's everyday lives has not kept up. On both fronts — safety and practicality — agents still appear to be in a transitional phase.

AMD to Acquire AI Chip Startup Taalas, Strengthening Its Data Center Offering

According to Bloomberg, AMD has reached a final agreement to acquire Taalas, a Canadian AI chip startup. The deal will reportedly allow AMD to offer a different kind of AI chip for the fast-growing data center market. Financial terms, including the purchase price, have not been disclosed.

There is further evidence of the strength of AI demand. Cloudflare raised its full-year profit forecast, beating market expectations, driven by accelerating demand for network services amid the rapid spread of AI. From semiconductors to network infrastructure, a virtuous cycle of investment and earnings is spreading across the entire stack that underpins AI.

xAI and SpaceX: The AI Infrastructure Construction Race Draws a Big Reaction

An article examining the "AI infrastructure construction race" surrounding Elon Musk's xAI and SpaceX drew a large response on Hacker News, picking up 113 points and 69 comments. The piece appears to focus on the data center building boom needed to sustain massive computing power, and the environmental strain that comes with it.

The technical details and figures require checking the original, but the sheer volume of community discussion is meaningful in itself. It is a sign that the physical infrastructure and energy issues underpinning AI's growth have already become a real concern for many engineers. Going forward, the focus is likely to extend beyond the race for model performance to the competition over construction and resources behind it.

OpenAI Researcher Departs to Develop AI That Reads Thoughts

According to India Today and other outlets, Naomi Bashkansky, formerly a researcher at OpenAI, has left the company to take on a new challenge: developing "AI that reads thoughts." It is an ambitious domain involving the interface between the brain and AI, and their future moves are sure to draw attention. No details on productization have been disclosed at this time.

MiniMax-H3 Generates Video in 4 Steps as the Community Releases a Speed-Up LoRA

Efforts to accelerate MiniMax-H3, a video generation model with synchronized audio, are being driven by the community. According to a repository that has been drawing attention on Hugging Face, using a "Turbo LoRA" for MiniMax-H3 can cut the generation of video and synchronized audio — normally requiring around 20 steps — down to roughly 4 steps. Sampling time is expected to shrink to about one-fifth.

The original Turbo LoRA was created and released by larryvrh and is currently considered a preview version that is still training. drbaph subsequently converted it into a version loadable in ComfyUI and also released a further-trained checkpoint (ckpt500). Users can now choose between EMA (a moving average of the weights) and non-EMA versions, along with the further-trained variants. Since the release of the H3 base model was reported last month, practical speed-ups and ecosystem formation have been advancing rapidly.