Thomas Joos writes that choosing LLM deployment now means picking between cloud APIs, datacenter GPUs or local workstations, with technical limits set by memory and bandwidth rather than raw compute and software choice tied to load profile. Quantization shrinks models to fit consumer GPUs and unified-memory systems, while token generation is bandwidth bound so bandwidth per dollar matters more than TFLOPS. Economically, self-hosting an eight-H100 cluster costs about $237k per year over three years versus higher on-demand cloud pricing, but break-even needs 50-83% sustained utilization that teams typically reach only 40-65%, and the strongest case for local inference is sovereignty and resilience not pure cost.
- Mistral Large 3 ships as a 675B-parameter mixture-of-experts with 41B active per token under Apache 2.0
- Nvidia DGX Spark offers 128 GB unified memory at $4,699 list price since Feb 23 2026
- vLLM generates 44x the tokens per second of llama.cpp at 64 concurrent users on H200 with Llama 3.1 8B
- DevTk finds self-hosting breaks even only at several billion tokens per month versus budget hosted open models
- Martin Kuppinger says sovereignty is not an end in itself; security and resilience come first and exit path matters more than immediate rebuild
- Mistral Large 3 ships as a 675B-parameter mixture-of-experts with 41B active per token under Apache 2.0
- Nvidia DGX Spark offers 128 GB unified memory at $4,699 list price since Feb 23 2026
- vLLM generates 44x the tokens per second of llama.cpp at 64 concurrent users on H200 with Llama 3.1 8B
- DevTk finds self-hosting breaks even only at several billion tokens per month versus budget hosted open models
- Martin Kuppinger says sovereignty is not an end in itself; security and resilience come first and exit path matters more than immediate rebuild
Databricks co-founder and CTO Matei Zaharia has been honored with the 2026 ACM Prize in Computing, recognizing his massive impact on big data through the creation of Spark. Zaharia, an associate professor at UC Berkeley, argues that Artificial General Intelligence (AGI) is already a reality, though it should not be judged by human standards. He warns of the security implications of AI agents that mimic human behavior and expresses optimism about AI's ability to transform research and engineering. By automating complex tasks like molecular simulation and data compilation, Zaharia believes AI will become a universal tool for understanding information.
This article presents the history and evolution of the APL programming language, from its early development to its current state.
This issue of ACM Interactions explores the crucial role of User Experience (UX) in ensuring responsible AI (RAI) design and use. Articles discuss the importance of UX in shaping meaningful and trustworthy AI systems, the potential of 'reuser experience' for sustainable design practices, and emerging trends in AI interaction design.
An analysis of Large Language Models' (LLMs) vulnerability to prompt injection attacks and potential risks when used in adversarial situations, like on the Internet. The author notes that, similar to the old phone system, LLMs are vulnerable to prompt injection attacks and other security risks due to the intertwining of data and control paths.
REFERENCES
1. Abelson. H.. and disessa, A.A. Turtle Geomelry: The Computer as a
Medium for Exploring Mathematics. MIT Press, Cambridge, Mass.,
1981.
2. Bolt, R.A. Spatial data-management. Rep., Dept. of Architecture.
MIT. Cambridge. Mass.. 1979.
3. diSessa. A.A. A principled design for an integrated computational
environment. Hum.-Compui. Interaction I, 1 (1985). l-47.
4. disessa. A.A. Notes on the future of programming: Breaking the
utility barrier. In User-Centered Systems Design, D. Norman and
S. Draper. Eds. Lawrence Erlbaum. Hillsdale, N.J., 1986.
5. Goldberg. A.. and Robson. D. Smalltalk-80: The Language and Its Implementafion. Addison-Wesley. Reading. Mass., 1983.
6. Papert, S. Mindstorms:
Acknowledgments. We gratefully acknowledge the
efforts of all those members of the Boxer Groups at
MIT and Berkeley who have helped to make Boxer
(almost) a reality. Special thanks to Michael
Eisenberg, Gregor Kiczales, Leigh Klotz, Ed Lay,
and Jeremy Roschelle.
1. Abelson. H.. and disessa, A.A. Turtle Geomelry: The Computer as a
Medium for Exploring Mathematics. MIT Press, Cambridge, Mass.,
1981.
2. Bolt, R.A. Spatial data-management. Rep., Dept. of Architecture.
MIT. Cambridge. Mass.. 1979.
3. diSessa. A.A. A principled design for an integrated computational
environment. Hum.-Compui. Interaction I, 1 (1985). l-47.
4. disessa. A.A. Notes on the future of programming: Breaking the
utility barrier. In User-Centered Systems Design, D. Norman and
S. Draper. Eds. Lawrence Erlbaum. Hillsdale, N.J., 1986.
5. Goldberg. A.. and Robson. D. Smalltalk-80: The Language and Its Implementafion. Addison-Wesley. Reading. Mass., 1983.
6. Papert, S. Mindstorms:
Acknowledgments. We gratefully acknowledge the
efforts of all those members of the Boxer Groups at
MIT and Berkeley who have helped to make Boxer
(almost) a reality. Special thanks to Michael
Eisenberg, Gregor Kiczales, Leigh Klotz, Ed Lay,
and Jeremy Roschelle.
