Zyphra Standard-Depth Startup Diligence Report
Zyphra could be attractive if AMD-optimized long-context inference plus open research converts into enterprise customers and favorable unit economics; the public record does not yet prove that conversion.
Zyphra Standard-Depth Startup Diligence Report
Track / diligence-required. Zyphra has strong public technical and strategic-partner signals for an early AI infrastructure/model company, but standard investment diligence should treat financials, customers, unit economics, cap table, legal controls and supplier commitments as unverified until primary documents are reviewed.
No audited financials, revenue by product/customer, gross margin, cash runway, backlog or AR schedules are public.
Diligence request: Obtain full financial statements, ARR bridge, customer schedule, cloud cost/margin model and board-approved forecast.
IBM Cloud, AMD silicon/ROCm and TensorWave infrastructure are central to the product roadmap and likely cost base.
Diligence request: Review supplier contracts, capacity, minimum commitments, failover plans and utilization/margin data.
No top-customer list, revenue concentration, churn, NRR/GRR, pipeline or named production customer evidence is public.
Diligence request: Conduct customer reference calls and review contracts, usage cohorts, pipeline and churn.
CB list, IBM and secondary profiles support a near-$1B valuation, but public funding totals and dates conflict.
Diligence request: Reconcile cap table, financing documents, proceeds, share price, liquidation preferences and side letters.
Public assets include voice cloning, EEG research-use disclaimers and unmoderated model warnings, plus agentic execution risk.
Diligence request: Review red-team reports, safety policies, AUP enforcement, data rights, consent controls and incident history.
Public Terms reference separate customer agreement, AUP and DPA, but those documents, SOC/security audits, privacy controls and insurance are not reviewed.
Diligence request: Request SOC 2/security packet, customer terms, DPA, AUP, privacy policy, insurance, litigation and regulatory schedules.
Zyphra competes with larger frontier labs, hyperscalers, inference providers and neoclouds; open assets may be replicable.
Diligence request: Obtain win/loss, independent benchmarks, customer switching analysis and defensibility/IP memo.
Inference is launched, while several cloud/agent/post-training/dedicated-capacity capabilities are roadmap or not backed by public customer metrics.
Diligence request: Review product roadmap, GA criteria, backlog, engineering milestones and customer commitments.
The public record supports Zyphra’s unicorn-screening status and a major IBM/AMD infrastructure relationship, but not revenue quality, cash runway, gross margin, cap table, debt-like commitments, or forecast reliability.
inconclusive confidence: low
No audited or management income statements, balance sheets, cash-flow statements, backlog, AR aging, or revenue by product/channel/geography were found in public sources.
| metric | public signal | evidence quality | missing primary documents | risk read |
|---|---|---|---|---|
| Revenue / ARR | No company disclosure; Latka estimates 2024 ARR $8.8M | Weak | Audited income statements, ARR bridge, customer invoices, deferred revenue | Valuation may rely on R&D/strategic scarcity rather than demonstrated revenue. |
| Gross margin / COGS | Product depends on AMD/IBM/TensorWave compute; no unit economics disclosed | Medium for cost-driver, weak for numbers | Inference margin by model/context, cloud bills, GPU utilization, customer pricing | Long-context MoE workloads may pressure margins. |
| Cash / burn | Large funding/compute partnership implied; no cash balance | Weak | Balance sheet, cash runway, capex commitments, cloud prepay | Capital intensity likely high. |
| Backlog / AR | No named backlog/customer revenue table | Weak | Backlog, AR aging, signed customer contracts | Commercial traction cannot be verified publicly. |
inconclusive confidence: low
No company projections were public. Public growth drivers are the cloud launch, Maia/agent roadmap, open-source funnel, and strategic compute access; cost and financing assumptions are not disclosed.
partially verified confidence: medium
The $1B unicorn headline is supported by CB Insights and IBM, with secondary profiles around $989M-$1B, but shares outstanding, investor rights, option pool, debt and other dilutive instruments are non-public.
| capital item | public evidence | not public | diligence implication |
|---|---|---|---|
| Valuation | CB list $1B; IBM Series A $1B; Forge $989.39M post-money | Term sheet, share price, liquidation preferences, valuation cap | Verify valuation mechanics and investor rights before accepting unicorn headline. |
| Total funding | Forge/EquityZen about $121M-$122M; CB financial profile visible total-raised field $11M; Latka $111.4M | Cap table, SAFE/note conversions, debt, warrants | Contradictory public totals require primary financing schedule. |
| Ownership | Select investors in CB list and profiles | Founders, employees, investors, option pool, warrants | Dilution and control cannot be assessed publicly. |
| Debt / commitments | IBM/AMD/TensorWave multi-year infrastructure relationships | Take-or-pay, prepaid cloud, equipment leases, minimum spend, debt covenants | Compute commitments could be economically debt-like. |
partially verified confidence: medium
Financing history is partially triangulated from CB, IBM and secondary-market profiles. Tax, accounting policies, revenue recognition, NOLs and deferred revenue are not public.
| date | event or round | amount or valuation | investors or counterparties | confidence | notes |
|---|---|---|---|---|---|
| 2021 / 2020 | Company formation window | Not public | Founders | Medium-low | Forge/EquityZen say 2021; Tracxn/AI Market Watch say 2020. |
| 2024-10-01 | Seed / seed-1 rows | Forge: ~$11.42M plus small seed-1; post-money not primary verified | Bison Ventures, Future Ventures, Gaia, Metaplanet, others per Forge | Medium | Secondary-market profile; reconcile with CB profile total-raised field. |
| 2025-06-06 to 2025-06-09 | Series A / A-1 / unicorn-list date joined | CB: $1B valuation; Forge: $100M Series A + $10.1M A-1 and $989.39M post-money | CB list: Bison Ventures, Future Ventures, Intel Capital; IBM says Series A closed at $1B valuation | Medium-high | Primary source for amount/cap table still required. |
| 2025-10-01 | IBM/AMD infrastructure collaboration announced | No purchase amount disclosed; IBM repeats $1B Series A valuation | IBM, AMD | High for partnership; low for economics | Multi-year cloud/GPU infrastructure, not necessarily equity financing. |
Zyphra’s public product surface is broad for its stage—open foundation models, speech/EEG research, MAIA agent, Inference, Compute and Agent Environments—but commercial maturity, pricing, production reliability and safety controls need primary validation.
partially verified confidence: medium
Public materials verify Zyphra Research, Zyphra Cloud, serverless Inference and a pipeline of agent/compute/environment capabilities. Public materials do not verify customer adoption, margins, pricing, market share or production SLAs.
| asset | category | disclosed status | target user or use case | evidence strength | key diligence question |
|---|---|---|---|---|---|
| Zyphra Inference | Commercial cloud service | Available / launched; serverless inference first component | Long-horizon agentic workloads, large MoE serving | Strong company evidence | Verify paying customers, pricing, uptime, margin, and SLAs. |
| MAIA / Agent | Agent product | Publicly positioned as general/multiplayer enterprise agent; adoption metrics not public | Enterprise knowledge workers and agentic workflows | Medium | Separate demo capability from production deployment and usage. |
| Agent Environments | Sandbox/runtime | Listed by cloud pages; some components described as upcoming | Isolated code, browser, terminal, and workflow execution | Medium | Verify security architecture, tenant isolation, SOC controls. |
| Compute / dedicated clusters | Infrastructure | GPU clusters, bare metal, dedicated capacity described; dedicated clusters/bare metal listed as upcoming in blog | Training, RL, simulations, custom AMD kernels | Medium | Verify supply commitments and resale/hosting economics. |
| Zonos | Open-weight TTS model | Public GitHub model with strong community interest | Text-to-speech, voice cloning, multilingual audio | Strong technical evidence | Assess voice-cloning safety, license obligations, and monetization path. |
| Zamba / Zamba2 / ZUNA | Open research models | Public blogs and repositories disclose architecture and limitations | Efficient language models, EEG foundation modeling | Strong technical evidence | Independently reproduce benchmark, data-rights, safety, and clinical-use claims. |
| dimension | what is public | what is missing | risk or opportunity |
|---|---|---|---|
| Pricing | Cloud/inference marketing pages and contact-sales flows; no full public rate card found | Token prices, reserved-capacity rates, discounts, committed-use contracts | Pricing opacity limits margin and adoption assessment. |
| Availability | Inference described as launched; future post-training/RL/environments/dedicated capacity described | Region availability, service limits, beta/GA definitions, uptime history | Roadmap risk remains material. |
| Safety/compliance | Terms reference AUP/DPA/customer agreement; demos should not receive sensitive or regulated data | AUP text, privacy/SOC packet, model cards for every hosted model, abuse-monitoring metrics | Enterprise procurement may stall without controls. |
| Differentiation | AMD/ROCm/custom-kernel/long-context positioning and open research portfolio | Independent TCO benchmarks versus Nvidia/cloud alternatives | Could be strong if verified; risky if company-authored benchmarks do not generalize. |
Public customer diligence is the weakest area: no top-customer list, revenue by customer, churn or concentration table is public. Disclosed relationships are strategic infrastructure partnerships rather than proven paying-customer adoption.
inconclusive confidence: low
No top-15 customers, product ownership, purchase timing or customer-level revenue were found in public sources. Public segments are developers, enterprises/knowledge workers and frontier AI hyperscalers.
| relationship or segment | type | public evidence | revenue evidence | diligence read |
|---|---|---|---|---|
| Developers | Target segment | Cloud pages target developers and agentic workflows | No revenue disclosed | Developer traction should be verified through usage cohorts and paid conversion. |
| Enterprises / knowledge workers | Target segment | MAIA described for enterprise knowledge workers and long-horizon workflows | No named enterprise logos | Obtain top-customer list and pipeline stage definitions. |
| Frontier AI hyperscalers | Target segment | About/Cloud pages name hyperscalers as target users | No contracts disclosed | Assess whether Zyphra is supplier, customer, competitor, or partner in each account. |
| Open-source users | Community demand signal | GitHub stars/forks on Zonos and other repos | No conversion data | Community adoption is useful but not equivalent to ARR. |
verified confidence: high
IBM Cloud, AMD and TensorWave relationships are well-supported publicly, but relationship economics, exclusivity and revenue/customer impact are not public.
| counterparty | relationship type | public terms | evidence strength | unknowns |
|---|---|---|---|---|
| IBM Cloud | Training infrastructure provider / cloud partner | Multi-year agreement; large AMD MI300X cluster; initial deployment in September 2025 and expansion planned in 2026 | Strong for existence | GPU count, price, minimums, SLAs, exclusivity, termination rights. |
| AMD | Silicon/platform partner | MI300X training, MI355X inference, AMD networking hardware; quotations in public releases | Strong for existence | Equity participation, product co-marketing economics, roadmap commitments. |
| TensorWave | Inference infrastructure provider | MI355X infrastructure provider for Zyphra Inference | Strong for existence | Capacity allocation, cloud spend, resiliency, resale margins. |
| Open-source ecosystem | Research distribution/community | Public GitHub repos and open model checkpoints | Strong for visibility | License compliance, contributor agreements, enterprise conversion. |
inconclusive confidence: low
No revenue by customer or any customer accounting for 5%+ of revenue is public.
inconclusive confidence: low
No severed customer, partner or supplier relationships surfaced in accessible public searches, but there is no public relationship ledger.
partially verified confidence: medium
AMD, IBM Cloud and TensorWave are highly visible suppliers/partners. Purchase amounts and supplier contracts are not public.
| supplier or dependency | criticality | evidence | purchase amount public | risk note |
|---|---|---|---|---|
| AMD Instinct MI300X / MI355X silicon and ROCm stack | High | Training and inference positioning are centered on AMD hardware/software | No | Roadmap, software maturity, and supply constraints could affect execution. |
| IBM Cloud | High | Large multi-year training cluster | No | Commitment terms may drive burn or constrain flexibility. |
| TensorWave | High for launched inference | Named MI355X infrastructure partner | No | Availability, utilization, and redundancy should be tested. |
| Open-source model ecosystems / third-party models | Medium | Inference serves external models such as Kimi, DeepSeek, GLM | No | Licensing, model availability, and supplier changes can affect service continuity. |
Zyphra has a differentiated public narrative—open models plus AMD-optimized long-context inference—but competes against heavily funded model labs, hyperscale clouds, inference specialists and neocloud providers with better-known customer scale.
partially verified confidence: medium
Zyphra’s basis of competition is technology architecture, open research, AMD optimization, long-context inference and enterprise agents. Public evidence does not prove market share, price leadership or customer preference.
| segment | representative competitors | zyphra positioning | public strength | public weakness |
|---|---|---|---|---|
| Frontier model labs / agents | OpenAI, Anthropic, Google DeepMind, xAI, Mistral, Cohere | Open superintelligence, MAIA agent, multimodal models, open research | Open-source technical footprint and strategic infrastructure | No disclosed customer/revenue scale versus much larger incumbents. |
| Inference/cloud model serving | Together, Fireworks, Replicate, Cerebras, Groq, hyperscale clouds | AMD-optimized long-context inference for large MoE/agentic workloads | Clear hardware differentiation and partner story | Pricing, uptime, independent benchmarks, and customers not public. |
| AI infrastructure / neocloud | Lambda, CoreWeave, Crusoe, TensorWave, cloud GPU providers | Product company plus compute layer rather than pure capacity rental | IBM/AMD cluster access and TensorWave relationship | Supplier concentration and unclear resale economics. |
| Open models / research communities | Meta Llama ecosystem, Mistral, Hugging Face model publishers | Open models across language, speech, and EEG with architectural experimentation | GitHub engagement and disclosed research assets | Open assets can diffuse moat unless monetization is differentiated. |
| basis | zyphra public score 1 to 5 | supporting evidence | main validation needed |
|---|---|---|---|
| Technology architecture | 4 | Zamba/Zamba2/Zonos/ZUNA repos and blogs disclose architectures and datasets | Independent benchmark reproduction, data lineage, safety review. |
| Cost/performance | 3 | AMD/ROCm/custom-kernel and MI355X memory-positioning claims | Workload-specific TCO versus Nvidia and other inference providers. |
| Distribution | 2 | Website, GitHub, blogs, contact sales; no channel partner revenue | Pipeline, win/loss, customer references, sales productivity. |
| Trust/compliance | 2 | Terms reference AUP/DPA/customer agreements but public compliance package is absent | SOC 2, security architecture, model safety governance, privacy review. |
| Capital access | 4 | $1B unicorn listing and strategic IBM/AMD relationship | Cash runway, future financing needs, obligations under compute deals. |
Zyphra’s public GTM engine is technical content, open-source distribution, product pages/contact-sales and strategic partner announcements. There is no public sales productivity, funnel conversion, pipeline or budget evidence.
partially verified confidence: medium
Public positioning emphasizes open/sovereign AI, long-horizon agents, AMD-optimized inference and enterprise/hyperscaler use cases. Implementation metrics are not public.
| channel | observable activity | buyer or user | public metric or gap | diligence request |
|---|---|---|---|---|
| Company website / cloud pages | Product positioning and contact-sales paths | Developers, enterprises, hyperscalers | No traffic, signup, activation, or paid conversion data | Provide funnel by source, activation and paid conversion by cohort. |
| GitHub/open-source | 23 public repos; Zonos has high stars/forks | Researchers and developers | Stars/forks visible; enterprise conversion unknown | Show repo-to-cloud conversion and enterprise leads. |
| Strategic partner PR | IBM/AMD/TensorWave announcements | Enterprise buyers and AI infrastructure buyers | Partner economics and co-sell activity not public | Provide signed co-marketing/co-sell terms and sourced pipeline. |
| Model launch/blog content | Inference, Zamba, model pages and research posts | AI builders, technical evaluators | No MQL/SQL attribution public | Provide attribution from content to opportunities. |
| public claim or signal | source | current status | verification needed |
|---|---|---|---|
| Inference built for long-horizon agentic workloads | Company pages/blogs | Publicly asserted | Production workload logs, latency/throughput by context length, customer references. |
| Open superintelligence and sovereign control | Homepage/About | Positioning verified | Security/privacy controls, deployment options, model customization, regulated-industry proof. |
| Enterprise superagent MAIA | IBM release/company cloud pages | Roadmap/product direction verified | Live deployments, active users, retention, accuracy and safety metrics. |
| $1B valuation and Series A | CB/IBM/secondary profiles | Headline verified, terms not public | Closing docs, cap table, investor list, option pool and preferences. |
inconclusive confidence: low
No major customer relationships or customer trend data are public; strategic partner relationships are not necessarily customers.
partially verified confidence: medium
Observable avenues are open-source/community, product-led website/cloud access, partner credibility and technical content.
inconclusive confidence: low
No sales headcount, quotas, compensation, productivity, sales-cycle or hiring plan data were public.
inconclusive confidence: low
Marketing budget and implementation capacity are not public. Compute commitments may compete with GTM budget for cash.
R&D is Zyphra’s strongest public diligence area: models, repos and technical blogs substantiate research activity. The open questions are cost, data rights, safety validation, key-person concentration, and conversion from research to durable commercial products.
partially verified confidence: medium
Public sources show active model research and named founders/leadership, but full org structure, personnel biographies, IP assignments and retention plan are not public.
| person or group | public role or signal | evidence | confidence | diligence need |
|---|---|---|---|---|
| Krithik Puthalath | CEO and Chairman; founder/CEO in partner releases/profiles | IBM release and company/market profiles | High for CEO/Chairman | Biography, employment agreement, equity, board role, technical involvement. |
| Tomás Figliolia | Cofounder in secondary profiles | Forge/Tracxn/AI Market Watch | Medium | Current role, reporting line, retention package. |
| Beren Millidge | Cofounder in secondary profiles | Forge/Tracxn/AI Market Watch | Medium | Current role, publications, IP assignment. |
| Danny Martinelli | Cofounder / products and partnerships; low-confidence sources conflict on CEO role | EquityZen/Forge/AI Market Watch/Latka | Medium-low | Resolve title, responsibilities, customer relationships. |
| Research engineering team | Public model work; Zamba blog claims seven-person team for model training | Zamba blog and GitHub | Medium | Org chart, hiring plan, publication/IP ownership, key-person risk. |
partially verified confidence: medium
The pipeline includes launched inference, MAIA, post-training/RL/fine-tuning, agent environments, dedicated clusters and open research models. Public materials identify several items as upcoming, so timing and development cost remain unverified.
| pipeline item | status from public sources | critical technology | development cost visibility | risk |
|---|---|---|---|---|
| Zyphra Inference | Launched/available as first cloud component | AMD MI355X, ROCm, custom kernels, long-context inference | No cost disclosed | Margin and uptime for long-context workloads. |
| Distributed post-training / RL / fine-tuning | Upcoming | GPU cluster, data pipelines, RL infrastructure | No cost disclosed | Roadmap slippage and compute burn. |
| Agent environments / sandboxed dev environments | Cloud page plus upcoming roadmap | Isolation, tool orchestration, secure execution | No cost disclosed | Security, abuse, tenant isolation. |
| MAIA general-purpose enterprise agent | Named in IBM release and cloud/product pages | Multimodal models, long-term memory, continual learning | No cost disclosed | Enterprise reliability and safety validation. |
| Open model research portfolio | GitHub/blog repositories for Zamba, Zamba2, Zonos, ZUNA | SSM/transformer hybrids, speech synthesis, EEG diffusion models | Mostly not disclosed | Data rights, license, safety, and monetization. |
Zyphra appears founder-led with Krithik Puthalath publicly identified as CEO/Chairman, but public headcount, roles, compensation, turnover and employment arrangements are not verified and commercial databases conflict materially.
partially verified confidence: medium
Only a public leadership sketch can be constructed. Internal reporting lines, board composition and functional headcount are not public.
inconclusive confidence: low
Commercial databases show conflicting team-size estimates ranging from about 20 to 65 and differing location/founding-year signals.
| source | founding year or stage | headcount or team size | location signal | confidence | note |
|---|---|---|---|---|---|
| Tracxn | Founded 2020; Series A | 65 as of Mar. 2026 | San Francisco | Low-medium | Useful for triangulation but funding date conflicts with CB/Forge. |
| AI Market Watch | Founded 2020 | 20 | San Francisco plus Montreal/London | Low | Conflicts with Tracxn/Latka; use only as directional. |
| Latka | 2024/2025 profile fields | 44 | Palo Alto | Low | Revenue and leadership fields conflict with stronger sources. |
| Forge/EquityZen | Founded 2021 | Not disclosed in rows reviewed | Palo Alto / San Francisco | Medium | Better for financing/location than for employee census. |
partially verified confidence: medium
Public profiles identify a founder/leadership group, but detailed biographies, ages, tenure, prior employment and current responsibilities are incomplete.
| name | public role | source confidence | material follow up |
|---|---|---|---|
| Krithik Puthalath | Founder; CEO and Chairman in IBM release | High | Full bio, employment agreement, board composition, key-person plan. |
| Danny Martinelli | Cofounder; products and partnerships in EquityZen; conflicting Latka CEO field | Medium-low | Confirm current title, customer ownership, retention and equity. |
| Beren Millidge | Cofounder in commercial profiles | Medium | Confirm current title, research/IP assignments. |
| Tomás Figliolia | Cofounder in commercial profiles | Medium | Confirm current title, reporting line, retention risk. |
inconclusive confidence: low
No employment agreements, founder compensation, severance, benefits or bonus terms were public.
inconclusive confidence: low
No option pool, equity incentive plan, grants, vesting schedules or exercises were public.
inconclusive confidence: low
No public employee-relations issue surfaced in accessible searches, but no HR/legal records were reviewed.
inconclusive confidence: low
Turnover data are not public and cannot be inferred reliably from database estimates.
No public red-flag litigation, SEC company match, IPO, acquisition or shutdown signal was found, but legal diligence remains incomplete. Key exposures are model safety, data rights, voice/biometric misuse, EEG/medical disclaimers, customer terms, supplier contracts and missing IP/insurance schedules.
inconclusive confidence: low
No direct lawsuit against Zyphra surfaced in accessible public searches, but comprehensive docket searches were not completed due public-source/tool limits.
| area | public result | confidence | diligence follow up |
|---|---|---|---|
| Pending lawsuits against Zyphra | No direct public lawsuit surfaced in accessible web search results reviewed | Low-medium | Run full litigation search across federal/state dockets, arbitration, demand letters, and counsel confirmation. |
| Lawsuits initiated by Zyphra | No direct public result surfaced in accessible web searches | Low-medium | Request litigation schedule and IP enforcement docket. |
| SEC public company status | EDGAR company search returned no matching companies | Medium | Confirm no confidential filing, Reg D filings, or entity-name variants. |
| Regulatory agency problems | No direct agency action surfaced in public sources reviewed | Low | Request regulatory inquiry/complaint schedule and compliance counsel memo. |
| Terms / governing law | Website terms updated May 4, 2026; California law/San Francisco courts; site liability cap $100 | High | Review customer agreement, AUP, DPA, privacy policy, order forms, SLAs. |
inconclusive confidence: low
No public lawsuit initiated by Zyphra surfaced in accessible searches.
partially verified confidence: medium
Physical environmental exposure appears low for a software/AI company, but data-center energy/carbon, employee safety, and AI safety/abuse risks require diligence. Public terms restrict misuse and warn against sensitive data in demos.
partially verified confidence: medium
Public GitHub/model assets and website terms are visible, but no complete IP schedule, patent/trademark inventory, data-license schedule or contributor agreement review was available.
| asset or right | public evidence | license or control public | risk |
|---|---|---|---|
| Zyphra website/materials | Terms claim Zyphra and licensors own site materials | Site evaluation/documentation use only; competing AI training restriction in Terms | Need customer-data/model-output IP terms. |
| Open-source repositories/models | 23 public GitHub repositories and model READMEs; Zamba blog references Apache 2.0 checkpoints | Varies by repository/model; full schedule not reviewed | Open licensing can reduce defensibility; compliance and contributor/IP assignment need review. |
| Speech/voice model Zonos | Open-weight TTS with voice cloning | Repository-specific license/usage controls require legal review | Deepfake, consent, biometric voice, copyright/right-of-publicity exposure. |
| EEG model ZUNA | Research-use-only; not medical/clinical use | Research disclaimer public | Medical-device, human-subject/data-rights, and health-data risks if commercialized. |
| Patents/trademarks | No patent/trademark schedule verified in this pass | Unknown | Freedom-to-operate and brand protection remain unverified. |
inconclusive confidence: low
No cyber, E&O, D&O, media/IP, product liability or workers compensation coverage details were public.
| contract or exposure | publicly known | undisclosed terms | risk level |
|---|---|---|---|
| IBM Cloud multi-year GPU cluster | Large AMD MI300X training cluster for Zyphra; expansion planned | Fees, minimums, uptime/SLA, data residency, termination, exclusivity, capacity | High |
| AMD collaboration / MI355X platform | AMD silicon central to cloud and inference story | Equity/investment terms, roadmap commitments, warranty/support, supply guarantees | High |
| TensorWave infrastructure | MI355X infrastructure provider for Zyphra Inference | Capacity, price, failover, shared responsibility, security | High |
| Customer agreements/AUP/DPA/orders | Terms say service use is governed by separate agreements | Customer data rights, indemnity, SLAs, model output ownership, limitations, audit rights | High |
| Insurance coverage | No insurance disclosure found | Cyber, E&O, D&O, media/IP, workers comp | Medium-high |
partially verified confidence: medium
Public materials identify IBM, AMD and TensorWave relationships and separate customer agreements, but signed contracts and material terms are not public.
partially verified confidence: medium
No SEC public-company match or direct regulatory agency problem surfaced, but regulatory risk remains from AI safety, privacy/data processing, voice cloning, EEG/health data, export controls and customer data handling.
| ID | Claim | Status | Sources |
|---|---|---|---|
| EC-001 | CB Insights unicorn list row identifies Zyphra as a $1B Enterprise Tech unicorn that joined on 6/9/2025, located in Palo Alto, United States, with Bison Ventures, Future Ventures, and Intel Capital as select investors. | verified high | SRC-001SRC-002 |
| EC-002 | CB Insights company financials page verifies Zyphra has a public profile with four funding events and latest Series A dated June 9, 2025, but the page’s visible total-raised value and gated valuation fields do not fully reconcile with the unicorn-list valuation. | partially verified medium | SRC-003SRC-001 |
| EC-003 | IBM/AMD announced a multi-year infrastructure collaboration for Zyphra to train frontier multimodal foundation models on AMD Instinct MI300X GPUs on IBM Cloud; the release says Zyphra recently closed Series A at a $1B valuation and identifies Krithik Puthalath as CEO and Chairman. | verified high | SRC-004 |
| EC-004 | Independent industry coverage corroborates the IBM/AMD GPU-cluster relationship while noting GPU count was not disclosed and an initial deployment was available in September 2025 with planned expansion in 2026. | verified medium | SRC-005SRC-025 |
| EC-005 | Secondary-market profiles present Zyphra as pre-IPO, founded around 2021, headquartered in Palo Alto/San Francisco, with approximately $121M-$122M total funding and a near-$1B Series A post-money valuation; these are secondary estimates, not company financial statements. | partially verified medium | SRC-006SRC-007 |
| EC-006 | Zyphra’s own site positions the company as building a full-stack for open superintelligence and emphasizes sovereign AI control, transparency, safety, alignment, open foundation models, and heterogeneous compute. | verified high | SRC-008SRC-009 |
| EC-007 | Zyphra Cloud is publicly presented as a four-part platform—Agent/MAIA, Agent Environments, Inference, and Compute—with emphasis on long-horizon agentic workflows, long context, model/tool orchestration, distributed training/RL, and AMD-optimized infrastructure. | verified high | SRC-010SRC-011 |
| EC-008 | Zyphra announced Zyphra Inference as a production-grade service for large MoE and long-running agentic workloads, powered by AMD MI355X GPUs in partnership with TensorWave and serving Kimi K2.6, DeepSeek V3.2, and GLM 5.1. | verified high | SRC-012SRC-013 |
| EC-009 | Zyphra’s AMD/TensorWave blog says the initial commercial service is serverless inference and that distributed post-training, RL/fine-tuning, sandboxed agent environments, dedicated clusters, and bare-metal capacity are upcoming rather than fully evidenced as generally available. | verified high | SRC-013 |
| EC-010 | Zyphra’s public model work includes the Zamba hybrid architecture, with company blog claims of open Apache 2.0 checkpoints trained by a seven-person team on 128 H100 GPUs in 30 days. | verified medium | SRC-014 |
| EC-011 | Zyphra has a visible public open-source footprint: 23 GitHub repositories were observed, led by Zonos, zuna, BlackMamba, Zamba2, and tree_attention by public stars/forks at research time. | verified high | SRC-015 |
| EC-012 | Zonos is an open-weight TTS model trained on more than 200k hours of multilingual speech with voice-cloning capability and high-quality audio output, which is a technical asset and a misuse/safety exposure. | verified high | SRC-016 |
| EC-013 | ZUNA is presented as a 380M-parameter EEG foundation model trained on about 2M channel-hours with explicit research-use-only and not-for-medical/clinical-use disclaimers. | verified high | SRC-018 |
| EC-014 | Zamba2 is presented as a 2.7B hybrid SSM/transformer model trained on 3T tokens plus 100B annealing tokens, but its README warns the base model is not instruction-tuned and may produce harmful or offensive content. | verified high | SRC-017 |
| EC-015 | Zyphra Terms of Use indicate commercial cloud/API users are governed by separate customer agreement, Acceptable Use Policy, Data Processing Agreement, and orders; public demos may log and process activity; benchmarks are informational only; site liability is capped at $100. | verified high | SRC-019 |
| EC-016 | Public leadership and headcount data are directionally consistent that Zyphra is founder-led by Krithik Puthalath with named cofounders, but commercial databases conflict on founding year, financing date, team size, and some leadership roles. | partially verified medium | SRC-004SRC-006SRC-007SRC-020SRC-021SRC-022 |
| EC-017 | Public revenue, ARR, gross margin, burn, cash, debt, backlog, and customer concentration are not verified; one low-confidence commercial estimate reports 2024 ARR of $8.8M and 44 employees, but it conflicts with stronger sources on leadership and funding. | inconclusive low | SRC-022SRC-003 |
| EC-018 | No public IPO, acquisition, shutdown, direct SEC company match, or directly surfaced lawsuit was found in accessible sources during this public-screening pass; this is not a legal opinion or exhaustive court search. | partially verified medium | SRC-001SRC-006SRC-008SRC-013SRC-023SRC-024 |
| EC-019 | No named paying customer list, revenue by customer, or customer accounting for 5%+ of revenue is publicly disclosed; the public demand signal is mostly strategic partner announcements, product pages, GitHub adoption, and target-segment language. | partially verified medium | SRC-009SRC-010SRC-012SRC-015 |
| EC-020 | Zyphra’s strategic dependencies are concentrated around AMD silicon, IBM Cloud MI300X training infrastructure, and TensorWave MI355X inference infrastructure; terms, exclusivity, minimum commitments, and take-or-pay economics are not public. | verified high | SRC-004SRC-005SRC-012SRC-013 |
| EC-021 | Zyphra competes in crowded frontier model, AI agent, inference, and AI infrastructure markets where public unicorn data already includes highly valued model labs and where differentiation depends on open models, AMD optimization, long-context inference, and enterprise agent workflows. | partially verified medium | SRC-001SRC-006SRC-010SRC-011SRC-012 |
| EC-022 | Commercial maturity remains early from public evidence: Inference is described as available/launching, while agent environments, post-training/RL, fine-tuning, dedicated clusters, and bare metal are disclosed as upcoming or not accompanied by customer metrics. | verified medium | SRC-010SRC-013 |
| EC-023 | Public legal/IP diligence is incomplete: open repositories and Terms are visible, but no patent/trademark schedule, insurance coverage, signed material contracts, SOC/security audit, or complete privacy/compliance packet was reviewed. | inconclusive medium | SRC-015SRC-019SRC-024 |
| EC-024 | Technical evidence is unusually visible for a young unicorn because public repositories, model READMEs, and company blogs disclose architectures, datasets, hardware, and limitations; however, benchmark and safety claims remain company-authored and require independent technical validation. | verified medium | SRC-014SRC-016SRC-017SRC-018SRC-019 |
This report is a public-evidence diligence snapshot, not investment advice. Important financial, legal, technical, and contractual facts remain non-public and should be verified directly with management and primary documents before any investment decision.