Rupen Chanda

Here’s my take on our newest blog: Industry Leaders Launch OpenAV Cloud for an Open, Interoperable AV Ecosystem AV leaders unite under OpenAV Cloud to tackle proprietary system challenges, introducing an open, interoperable framework powered by secure, cloud-based APIs. This initiative is a game-changer for flexible and customer-first innovation. Read the full blog for more details: https://hubs.li/Q03qsYKC0 #AVInteroperability #UnifiedAVEcosystem #AVIntegrationSolutions https://hubs.li/Q03qsYKC0

17

1 Comment

The $500 billion question nobody's asking about AI data centers. Stargate: up to 500billion. Microsoft: 80 billion. OpenAI pays Nvidia to get paid by Nvidia. AMD offers OpenAI 10% equity if they buy enough GPUs. These aren't investments anymore—they're circular financial instruments disguised as infrastructure. The downstream effects are already here. AI energy demand is set to surpass bitcoin mining this year. Meta's $27 billion Louisiana data center triggered a 600% spike in local vehicle crashes. Municipal water supplies are being drained for cooling systems. Here's what makes this fascinating from an investment perspective: demand justification is entirely circular. AMD's Lisa Su cited "overwhelming demand" from 800 million ChatGPT users. But consumer usage doesn't require gigawatt-scale data centers—enterprise deployment does. And enterprise is being funded by the same circular investments building the infrastructure. Rome didn't fall because the empire was a bad idea. It fell because the expansion math eventually stopped working. For those tracking AI investments: What's your thesis on when data center ROI actually materializes? What does rational scaling look like versus what we're seeing now? https://lnkd.in/e-hVbCtg #DataCenterEconomics #AIInfrastructure #TechCapitalDeployment

18

4 Comments

#genAI is changing password security. Tools like PassGAN can crack 51% of common passwords in under a minute by recognizing patterns humans create. Traditional complexity rules aren’t enough anymore—attackers now have sophisticated AI that predicts likely password combinations. Time to rethink password policies and consider breach monitoring alongside length-based requirements. #CyberSecurity #AI #threatmodeling https://lnkd.in/eQPaUnit

34

If GPU's could do this to CPU's, guess what QPU's could do... are you ready for this? 2025 to 2030 isnt about the GPU vs CPU narrative: The Roadmap to the Future Just a few years ago, CPUs were the backbone of all computing. Then came GPUs—engineered for massive parallelism, powering everything from gaming to AI. In 2025, the pace is accelerating: 1) AMD's MI350X & MI355X are shipping now, delivering nearly 4x performance year-on-year for AI and data workloads. The roadmap includes the MI400 and MI500 series, plus the rack-scale Helios AI system for 2026/27. 2) NVIDIA is pushing annual capacity increases and new architectures, with next-gen GPUs and advanced networking, targeting ever-larger AI clusters. 3) Gaming & Professional GPUs like the Radeon RX 9060 XT and AI PRO R9700 are setting new records in performance and AI capabilities for creators and professionals. 4) Quantifying the Difference: 4a. CPU vs GPU Metric CPU (2025) GPU (2025) 4b. Cores 8–96 5,000–20,000+ 4c. Parallel Tasks Limited Massive (thousands) 4d. AI Performance ~1 TFLOP 100–1,000+ TFLOPS 4e. Memory Bandwidth ~100 GB/s 1–8 TB/s 4f. Typical Use General compute AI, graphics, big data. Introducing QPU: The Next Leap Quantum Processing Units (QPUs) are emerging as the next revolution. While GPUs shattered CPU limits with parallelism, QPUs promise to solve problems impossible for classical computers—like breaking encryption, simulating molecules for drug discovery, and optimizing global logistics. If GPUs could do this to CPUs, just imagine what QPUs will do to everything we know about computing. Are you ready for the quantum era? #AI #QuantumComputing #GPU #QPU #FutureTech #Innovation Listing few startups who would give the biggies a run for their money: Cerebras Systems : Specializes in wafer-scale chips, with its latest WSE-3 featuring 4 trillion transistors and 900,000 AI cores. d-Matrix: Pursues in-memory compute architectures, aiming to overcome the traditional bottleneck between memory and compute for AI workloads. Rebellions : A South Korean startup focused on large language model (LLM) inference. Tenstorrent : Develops the Wormhole chip and related hardware for AI research and enterprise servers. Etched : Claims to have created the world’s first transformer ASIC, “Sohu,” designed specifically for LLM inference. Graphcore : British company known for its IPU-POD256 AI chip, recently acquired by SoftBank. Graphcore focuses on research and enterprise AI. Mythic : Innovates with analog compute architectures for edge AI, aiming for high power efficiency in devices like the M1076 AMP. Speedata.io : Based in Tel Aviv, develops Analytics Processing Units (APUs) for accelerating big data analytics and AI workloads. https://lnkd.in/dntV7DRC The Leonardo demo 80 millisecond demo 3 min video: https://lnkd.in/dvqmjjcc

19

1 Comment

Big news from the Gray Line portfolio — I’m thrilled to announce our lead investment in Archive Intel's $6.3M Series A! Compliance has historically been a tax on innovation. Archive Intel flips that script. Their AI-powered platform consolidates archiving, marketing review, and supervision — cutting manual review by up to 85% and reducing false flags by 99%. More than 500 wealth management and financial services firms and 7,000 users are already on board. Under Larry Shumbres's leadership, the team is redefining how regulated firms operate — not by layering on more tools, but by finally unifying compliance into one intelligent system. We’re proud to partner alongside Social Leverage and Garuda Ventures as Archive Intel scales into the category leader it’s clearly on track to become. https://lnkd.in/gqsTPQhA

158

8 Comments

For decades, engineers mistook a static curve for a universal truth. A measurement mic at the mix seat captured steady-state amplitude response; a “room curve.” The trace became law. Vendors built corrections around it. Systems sold on a promise that matching a target slope could fix loudspeaker power response, boundary gain & modal coloration in 1 step. It couldn’t. –– Professional practice moved on. Pink-noise RTAs gave way to time-windowed FFTs that isolate direct arrivals and early reflections within the impulse response. Alignment now means flat, smooth direct sound – neutral on-axis magnitude & phase, stable group delay – not a pretty trace. A mix engineer’s monitors, the dubbing stage, and the home loudspeaker all share a transfer function in one psychoacoustic loop. –– Every enclosure tells the same spectral story. At long wavelengths, radiation is near-omnidirectional (directivity index ≈ 0 dB). As frequency rises, beamwidth narrows and DI climbs beyond 10 dB. The listener receives a field weighted toward reflected energy below the transition region and direct energy above it. The upward tilt in in-room power response is not a defect, it is boundary-reinforced acoustics. –– Measurements from concert halls to cinemas confirm this. Heavy, high-STC walls trap bass energy; upholstered seating yields absorption coefficients near zero below 125 Hz. Most surface treatment dissipates only the mids and highs. The result is an RT60 that lengthens toward the low band – more modal density, longer decay, elevated steady-state level. –– Above the Schroeder frequency, the loudspeaker dominates; below it, the room does. A flat anechoic axial magnitude response defines timbre and precedence cues. Off-axis response defines spatial impression and warmth through total radiated sound power. Equalizing a steady-state trace without directional context is error: the microphone integrates scalar amplitude; the ear interprets vector fields. –– Equalization cannot alter directional radiation, reverberant field build-up, or non-minimum-phase interference. It cannot remove comb-filter notches that exist only at discrete points. It can only attenuate low-frequency resonant modes when spatially averaged and applied with psychoacoustic weighting. –– Below the transition region, parametric correction and subwoofer alignment can be beneficial; above it, filtering degrades coherence. A filter that flattens a measured curve often destroys the on-axis neutrality that defines accurate loudspeakers. Human adaptation tolerates reflections; it does not tolerate corrupted direct sound. –– The remedy is not “room correction, i.e EQ.” It is calibration grounded in acoustics and perception. Measure transfer function, directivity, decay time, and modal response, then correct where physics allows and hearing demands. Calibrate to the ears + brain system, not to a line on a plot. Image courtesy of Toole, adapted from: DOI: https://lnkd.in/gyb8MYkj.

22

1 Comment