5 startups that caught VCs’ attention at the latest PearX demo day

PearX, Pear VC’s 12-week accelerator, featured 16 startups at its latest San Francisco demo day, with five AI-focused companies drawing attention from investors. Saia claims its inference chip can deliver eight times the capacity while using four times less power than Nvidia’s Jetson; it plans test-chip fabrication next year and targets mass production by 2028. AI estate-planning startup Veros says it already manages $250 million in AUM and is seeking a trust charter.

Analysis

The investable signal is not the demo-day buzz; it is a set of hypotheses that could eventually pressure incumbent AI economics. Saia’s proposed flash-based inference architecture, if independently validated, could reduce power and memory constraints in edge devices. That would broaden deployment and potentially shift some edge workloads away from NVIDIA, but it is not evidence of near-term displacement: prototype performance, software compatibility, yields, and production economics remain unproven. The reported comparison with Jetson should be treated as a company claim, not an apples-to-apples benchmark. A successful design could also create a new integration opportunity for Samsung, but that is contingent on a working chip and commercial terms.

Speridlabs’ claimed editability points to a product distinction in spatial models, not proof of a durable moat. Alphabet’s Genie-related exposure is therefore a watch item, not a basis for changing the GOOG thesis absent evidence of customer adoption or monetization. For both Alphabet and NVIDIA, these startups are immaterial today relative to consolidated businesses; headlines risk overstating competitive urgency.

Over 1–3 months, the likely catalyst is follow-on financing or independent technical validation, not revenue impact. Over 6–18 months, monitor chip tape-out/test results, power-performance benchmarks, software ecosystem support, and customer pilots. The contrarian read is that accelerator scarcity can motivate architectural alternatives, but memory workarounds may trade bandwidth, latency, or endurance for capacity and power savings. A failed benchmark or delayed fabrication would weaken the disruption narrative. Other ventures described have regulatory, adoption, and execution hurdles; reported AUM or investor interest alone does not establish recurring revenue or defensibility.

Key Decisions for Investors

• No immediate position in GOOG or NVDA based on this coverage: the companies are private, the technical claims are unverified, and any plausible competitive impact is long-dated and small relative to the incumbents’ current businesses.
• Add Saia to an edge-AI watchlist. Reassess only after independent, workload-matched benchmarks and test-chip results; specifically verify latency, throughput, power, flash endurance, software support, and manufacturing yield before treating the Jetson comparison as economically meaningful.
• Track Speridlabs’ customer pilots and evidence that persistent, editable 3D outputs outperform alternatives in production workflows. Without adoption or monetization data, do not infer a material threat to Alphabet or other spatial-model providers.
• Treat Samsung integration, trust-company charter progress, and the startups’ fundraising as verification items—not catalysts by themselves. A delayed tape-out, weak benchmark, absent customer pilots, or regulatory setback would falsify the relevant upside claims.