AI inference stack — Daxos screen

The screen

The Lamb Labs teardown produced a specification, and we used it as the filter. Rides Nvidia's roadmap instead of racing it — a 12-month performance-per-watt reset kills any 18 to 24 month silicon cycle. No fab in the critical path — $50M is the NRE floor for a simple accelerator ASIC and only 14% of ASIC projects reach working first silicon. Batch-of-one or power-constrained, where the efficiency techniques genuinely win instead of inverting under load. Founders whose credentials survive primary-source checking. Under $15M raised.

1,442 companies came out of a 45-query Harmonic sweep. 77 passed the funding, age and headcount filters. Eight got full teardowns with an adversarial pass. Every funding, headcount and stage figure on this page is Harmonic.

The verdict

Rating	Company	Raised	HC	Where	What they are	What stops it
6.5	Heronic heronic.ai	$0	5	London	FPGA inference toolflow (MOSAIC) + a 4K super-resolution IP core. ASIC is a compiler target, not a plan.	Nobody has verifiably paid them. Chain of title to the core IP is unresolved.
5.5	Kernelize kernelize.ai	$2.0M	6	SF	Triton compiler backends for silicon the mainstream stack ignores. Best team we found.	Open Core Ventures owns the majority and the only board seat. No entry point.
5.5	CLIKA clika.io	$1.1M	12	San Jose / Seoul	Multi-backend compression SDK. In-Q-Tel on the cap table, verified active.	IQT is filed under Asia-Pacific with no US federal award. Engineering headcount 5→2 in a year.
5.0	ZETIC zetic.ai	$420K	11	Palo Alto / Seoul	Compiles models across Apple ANE, Qualcomm Hexagon, MediaTek. Publishes real per-device latency.	The $420K is accelerator money, not a round. Co-founder left Oct 2025. No watts published.
5.0	Sentinel Devices sentineldevices.com	$0 equity	5	Knoxville TN	Air-gapped edge ML anomaly detection for industrial control systems. $1.56M in verified federal awards.	Five people, never raised. May not want venture money at all.
4.5	myrtle.ai myrtle.ai	$11.1M	8	Cambridge UK	VOLLO FPGA accelerator, shipping on named parts with reproducible batch-1 microsecond latency.	$11.1M kills the funding bonus. Founded 2003 and still 8 people. No named customer.
3.5	Type 1 Compute type1compute.com	$200K	2	Palo Alto	On paper, the exact business the teardown specified: models onto FPGAs already fielded, defense edge.	Four third-party research works narrated in the first person. A headline metric with no source.
1.0	SqueezeBits squeezebits.com	$2.65M	21	Seoul	Quantization toolkit, Samsung/NAVER/Kakao backed, strong verified founder.	Acquired. A chip vendor bought them. Harmonic does not reflect it.

Lamb Labs, for calibration, sits at 3.0. Prior anchors: Apyx 9.0, Askari 8.5, BESXAR and Sophia 7.5, Katomed 7.0, Wootz 6.5, TeroAI 5.5, Auron 2.5.

6.5

Heronic Technologies

$0 raised · 5 people · London · incorporated 23 June 2024

This is the precise inverse of Lamb Labs, and it is the closest thing we found to the business the teardown said was underwritable. Where Lamb's founder forked a 179-star open-source accelerator the day before demoing it as his own, Alexander Montgomerie-Corcoran has 14 peer-reviewed papers in FPGA CNN mapping, original non-fork repos going back to 2020, and three independent third-party validations that survived hand-checking:

An MLPerf Tiny v1.1 closed-division submission in the MLCommons repo — the strict, rules-bound division, four boards, real numbers. Verified at mlcommons/tiny_results_v1.1/closed/fpgaconvnet/
First place in AMD's 2023 Open Hardware PhD category, credited on openhw.eu
Altera published a joint benchmark document and filmed a 4K demo on Altera's own YouTube channel. A silicon vendor does not do that for nobody.
$0 raised is real, not a Harmonic gap — confirmed against Companies House filings

What holds it at 6.5 rather than higher, and these are serious:

No verified paying customer. Every logo is a programme, a competition or a partner directory. Year-one accounts show £12,968 net assets and an average of zero employees — operationally this is a one-year-old company, not two.
The chain of title is unresolved. fpgaConvNet is Venieris and Bouganis's work, not his; he built the extensions on top. The public code sits on his personal GitHub under GPL-3.0 with four outside contributors and no CLA. Imperial holds no equity and Heronic is absent from Imperial's spinout register — yet a UK government delivery partner calls it "an Imperial College London spinout." One of those is wrong.
The team is thinner than the team page. The original CTO left for AMD. The current CTO's GitHub profile, updated two days ago, still names another employer.
For a super-resolution product they publish no image-quality metric at all — no PSNR, no SSIM, nothing. Quality is the product.

A clean written answer on the Imperial IP position plus one paying customer moves this to 7.5. A discovered Imperial claim, or GPL contamination of the commercial toolflow, drops it to 4.

5.5

Kernelize

$2.0M · 6 people · San Francisco · sole investor: Open Core Ventures

The team is a 7.5. The structure takes two points off. The résumé checks did not merely survive, they over-delivered from sources that cannot be faked:

"Simon Waters, Samsung" appears in the Khronos SPV_KHR_integer_dot_product contributor list — verifying the Khronos claim, the Samsung employer, the right years and exactly the right subject area in one artifact. Integer dot product is the INT8 quantized-inference primitive.
He owns AMD's Triton stream-pipeliner v2 upstream, has merged MLIR loop-pipelining patches in llvm/llvm-project, and is a CODEOWNER of triton-lang/triton-ext, a repo in the official Triton org.
The Tenstorrent relationship is verified in Tenstorrent's own repo — 18 merged PRs, continuous through July 2026. Not a logo.
Waters and Bowyer are co-inventors on Mentor Graphics patents filed in 2006. They have shipped compilers together for twenty years.
The three product repos are original, not forks, with hundreds of commits and pushes on the day of review. It passes the Lamb test decisively.

Open Core Ventures formed the company, holds the majority and the single board seat. Per OCV's own public handbook, founding CTOs get 10–25%, there is no CEO yet, and board approval is required for any contract over $50K. A $100–500K check buys 1–5% and zero governance.
Nothing is open. Harmonic currently_raising: null, and the round detail array is empty — it is an OCV SAFE, not a priced round.
Their own pricing page sells a $100–200K, one-to-two-month "Jumpstart" engagement as the entry product. First revenue is consulting. No signed platform licence is evidenced.
They gave the plugin mechanism away upstream, free and default-enabled in Triton 3.7+. Excellent for the ecosystem, ambiguous for the business.

Track, do not write. Revisit when OCV steps down at an external seed and a CEO is in place. This is the one to keep a calendar reminder on.

The defense lane is where the cleanest Daxos-shaped deals are

Two companies came out of the defense screen with the profile we usually want and rarely find: near-zero equity, real verified government revenue, and a first outside check still available. Neither got a full teardown in this pass and both deserve one.

Company	Equity	Verified federal awards	Founder
Sentinel Devices Knoxville TN · 5 people	$0 never raised	$1,556,218 — USAF $1,238,485 for a zero-cloud ICS analytics platform, preceded by a $68,714 award on the same topic (a Phase I to Phase II progression, which is the signal that matters), plus NRC $249,019 for nuclear cybersecurity	Forrest Shriver — DOE Innovation Crossroads Fellow at Oak Ridge National Laboratory, 2021–23
Tercero Technologies Arlington VA · 2 people	$100K	$2,810,236 all Navy, including an SBIR Phase II of $995,691 in March 2025. A 28:1 non-dilutive-to-equity ratio	Carl P. Evans III — Foster-Miller (TALON robots), QinetiQ, Neya Systems; sits on the board of the National Advanced Mobility Consortium, the OTA vehicle Army ground robotics actually flows through

The honest counter on both: two and five people respectively, and the classic SBIR-mill risk. The question for each is whether anything has ever sold outside a Phase II.

What we rejected, and the patterns worth keeping

The rejections are the more useful half of this exercise. Four checks did nearly all the work.

The is_current_position check killed six candidates in one round. Synseis — both founders hold full-time jobs elsewhere. Krakatoa — the CEO has left. RapidStream — acquired by a proprietary trading firm, both founders out. Acceleration Robotics — the one technical founder gone. Broccoli — founder consulting elsewhere. SpiceEngine — founder is a current RIKEN researcher. Run it first, before reading anything else.

Two companies selling "silicon IP" publish no HDL of any kind. OpenMachine sells silicon IP with zero Verilog, VHDL, Chisel or HLS files anywhere public — its 220-star repo is primarily TeX. Chameleon Semiconductor claims to be the only US soft-eFPGA IP provider and its sole founder is a career sales executive. Both fail the "is the RTL theirs" test one step earlier than the fork check: there is no RTL to compare.

Star counts routinely belong to something else. Wafer's headline 1,282-star repo is a reading list; its actual product repos have 16–21 stars. RightNow's 18,076 stars belong to an unrelated agent framework, while its in-lane repo was created and abandoned inside eight days. ggml is not investable at all — Georgi Gerganov's CEO tenure ended February 2026 and he is now a Hugging Face engineer.

deepsilicon is the cautionary datapoint. Harmonic stage: OUT_OF_BUSINESS. Ternary quantization plus custom silicon on a $500K seed — the Lamb Labs thesis, run to its conclusion.

Type 1 Compute deserves its own note

On paper it was the single best match to the specification: converts trained models onto FPGAs already fielded, no retraining, defense and autonomy, US entity, $200K raised. It rates 3.5 — below Lamb Labs on integrity, above it on substance.

Four separate third-party research works are narrated in the first person across four published documents — a Radboud University spiking RNN, BICLab's SpikeYOLO, and a TU Delft MAVLab result that ran on Intel Loihi, all uncited or credited as "we designed", "we achieved".
The headline 75.76 GOP/s/W has no traceable source in the literature. It is their own 41.67 divided by their own 0.55.
The Jetson baseline they compare against takes four different values across their own materials: 9.00, 8.00, 8.0 and 3.6.
The power comparison is a fabric-only chip estimate against a whole-board measurement.
Antler is claimed; a crawl of all 1,039 companies in Antler's portfolio does not list them. Stanford StartX Spring 2026 is verified.

Genuinely real: seven installable PyPI packages, a documented IR and graph partitioner, and one researcher whose published specialty matches the code. Pass, but ask the five questions — one Vivado utilization and power report for one model on one board would change the conversation.

What this says about the space

The honest headline is that we found nothing to write today, and the reasons are structural rather than bad luck. The best technology in the pool has no customers. The best team is majority-owned by its incubator. The best commercial traction is a Korean company that already got acquired. The cleanest government revenue sits in two companies of two and five people that may not want venture money.

That is not a null result. It is a map. The lane that survives contact — software and IP on commodity silicon, batch-of-one, no fab — is real and it is thinly populated, which is why the same six or seven names kept resurfacing across six independent sourcing agents. Heronic and Kernelize are both worth a call now even though neither is writable today, and the two defense companies are worth a proper teardown next.